Results in Engineering 24 (2024) 103553

Contents lists available at ScienceDirect

Results in Engineering
journal homepage: [Link]/journal/results-in-engineering

Review article

Computational methods for automatic traffic signs recognition in

autonomous driving on road: A systematic review
Hui Chen a , Mohammed A.H. Ali a,* , Yusoff Nukman a , Bushroa Abd Razak a , Sherzod Turaev b,
YiHan Chen c , Shikai Zhang a , Zhiwei Huang a , Zhenya Wang d, Rawad Abdulghafor e
a
Department of Mechanical Engineering, Faculty of Engineering, University of Malaya, 50603 Kuala Lumpur, Malaysia
b
Department of Computer Science and Software Engineering, College of Information Technology, United Arab Emirates University, Al Ain 15551, United Arab Emirates
c
Graduate School of Business, Universiti Sains Malaysia, 11800 Pulau Pinang, Malaysia
d
Department of Mechanical Engineering, Tsinghua University, 100084 Beijing, China
e
Faculty of Computer Studies (FCS), Arab Open University–Oman, Muscat P.O. Box 1596, Oman

A R T I C L E I N F O A B S T R A C T

Keywords: This review discusses the progress made in the traffic-sign detection and recognition methods and algorithms
Traffic sign detection (TSD) over the last decade with analyzing the strengths and drawbacks of each algorithm. The recent development of
Traffic sign recognition(TSR) traffic sign recognition on the roads highlights the necessity for precise detection of road’s traffic signs in various
Traditional-based sign detection algorithm
driving scenarios. In addition, the connections between the detection algorithms before and after the advent of
Deep Learning based sign detection algorithms
Hybrid algorithm-based sign detection
deep learning are revealed. The Traffic sign recognition has been developed to identify various shapes, sizes,
algorithms orientations, and appearances of signs in diverse conditions. Researchers have proposed numerous algorithms to
address these challenges. The traffic recognition methods have been categorized in this paper into three main
techniques, namely, conventional, deep learning, and hybrid based methods. The algorithms are compared with
each others via regression, segmentation, and hybrid techniques, specifically SSD, YOLO, Faster R-CNN, Pixel
Aggregation Network, and Mask R-CNN. The results demonstrate that the hybrid based detection algorithms
outperform others in true-positive rates, false-positive rates, the number of test images, accuracy, and processing
time. Such outcomes illustrate the potential of hybrid methods in the creation of accurate and effective TSD
systems, thereby paving the way for further research in this field.

Introduction signs using monocular cameras under diverse environmental conditions,

studies such as those by Li et al. and Hou et al. have investigated
Automatic traffic sign recognition involves the extraction of the improved methodologies for recognizing distracted driving contexts and
useful features and localization of the traffic signs in the designated live the implications of augmented reality (AR) for road cooperation in
streaming videos or images. Over the past two decades, more than 1800 traffic sign detection using in-car cameras [12–15]. These advantages of
research papers have been published in this domain, as shown in Fig. 1, in-car cameras for traffic sign detection have further underlined the
with several ground-breaking studies [1–11]. The rapid advancements importance of enhancing the adaptability of TSD systems to support
in deep learning and hardware capabilities for processing large volumes complex real-world [Link] continue to focus on
of image data have further propelled the progress in this field. Recent developing algorithms capable of processing real-time images in com­
advancements in-car camera systems and on-board computing resources plex real-world environments, such as those characterized by intense
have significantly influenced the ability of TSD systems to function light, rain, darkness, and blur. Another critical challenge in the devel­
effectively in real-world scenarios. Several studies have explored the opment of Traffic Sign Detection (TSD) systems is the variability of
usage of in-car cameras for real-time traffic sign detection, highlighting traffic signs across different countries. Traffic signs vary significantly in
the practical challenges of integrating such systems into vehicles. For terms of color schemes, symbols, shapes, and even language, depending
instance, the research by Lin et al. utilized a two-stage fusion neural on regional and national standards. For example, a "Stop" sign might be
network, demonstrating effective performance in recognizing traffic represented by different languages or slightly different shapes in various

* Corresponding author.
E-mail address: hashem@[Link] (M.A.H. Ali).

[Link]
Received 4 September 2024; Received in revised form 24 November 2024; Accepted 26 November 2024
Available online 7 December 2024
2590-1230/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license ([Link]
H. Chen et al. Results in Engineering 24 (2024) 103553

• Both classical and recently developed methods for traffic sign

recognition have been considered in this review to show the timeline
of the TSD development.

Fig. 2 shows the procedure of selection for WOS papers in this paper.
This review discusses the progress of TSD, including the traditional
TSD methods, deep learning algorithms, and hybrid approaches (Hybrid
algorithms refer to approaches that integrate both classical machine
learning techniques (e.g., feature-based detection or color segmenta­
tion) and deep learning methods to leverage the strengths of each. The
goal is to capitalize on the lower computational demands of classical
methods, along with the superior accuracy and robustness of deep
learning approaches. In some instances, hybrid methods may also
involve the combination of multiple deep learning strategies -such as
using data augmentation to increase dataset diversity or transfer
learning to leverage pre-trained models - to overcome challenges like
limited training data, lack of generalizability, and the need for enhanced
robustness against varying conditions.).The TSD has evolved through a
process that begins with preprocessing, followed by color segmentation,
shape recognition, feature extraction, and pattern matching before
Fig. 1. Publications on traffic sign detection published between 2000 and 2024 reaching the final stage of post-processing. Preprocessing is used to
as recorded in the WOS core collection.
prepare the image with its feature extraction, while post-processing is
aimed to refine the results by eliminating non-sign objects or improving
countries. Additionally, some countries use unique symbols to indicate classification. The Early published works are facing challenges with
specific actions, which may be unfamiliar to a detection model trained imbalanced datasets and limitations in deep learning [22]. However,
on signs from a different country. These differences present significant researchers have proposed alternative solutions over time, resulting in
obstacles for creating a universal TSD model that can adapt to the traffic more robust traffic sign detection algorithms that can handle complex
regulations of multiple regions, making it crucial to develop systems that conditions such as varying orientations and diverse sign appearances.
are versatile and capable of generalizing across diverse datasets. The Performance evaluation of traffic sign detection algorithms considers
goal is to create robust and adaptive TSD systems that can operate both accuracy and detection speed, especially for deep learning models.
effectively under the effect of various conditions. Two main review Efficient algorithms for road safety and traffic management are crucial
works introduced by [16,17] have reviewed the research advancements due to their applications in the advanced driver-based assistance sys­
on the road traffic sign recognition until 2019. At that stage, a few re­ tems (ADAS), self-driving cars, traffic monitoring, and infrastructure
searchers have utilized the deep learning strategy for detecting the planning. While many reviews have explored similar topics, the swift
traffic signs of the roads. On other hand, authors Liu, Tabernik, and Tian evolution of TSD algorithms, combined with the increasing drive to
have conveyed the recent advancement in the traffic sign detection commercialize autonomous driving technologies in recent years, re­
using the deep learning paradigm [2,18,19]. In fact, The recent devel­ quires a fresh examination of the latest research advances. As a result,
opment in computer hardware processing and software leads to a great this paper outlines its distinct contributions as follows:
enhancement in deep learning, especially for traffic sign detection which
is not reported yet and need a further review efforts to highlight the
current progress without overlooking to the foundational developments
in early stage of the traffic sign image detection like color, shaped based
detection [16]. Although the deep learning demonstrates a good efficacy
in traffic sign detection, it requires vast datasets for training the neural
networks and high computational resources to perform them which
represent as the main barriers for utilizing such AI in road sign detection
[2]. In addition, the training of the features in deep learning models
remains challenging as it depends on the data labelling process and
possesses a difficulty to get a ground truth data [2]. Consequently, the
integration of early-stage strategies with deep learning techniques
deems to be a promising method for developing an efficient road sign
detection model, which is the target of this review.
WOS journals provides a bibliometric information on the ISI pub­
lished papers and their citations [20,21], which is used to study the
progress of TSD system. Fig. 1 depicts the bibliometric data for TSD
works that have been published in the latest 24 years (2000–2024). The
papers selected for review on TSD have been selected based on the
following criteria:
Priority is given to impactful research and extensively refined
methodologies with well-cited and commonly applied techniques pa­
pers, rather than those with fewer than ten citations or lack of
recognition.

• The selected papers must demonstrate successful simulations and

experimental work with substantial outcomes in order to increase the
practicality and reliability of the study. Fig. 2. Flowchart for the paper select process.

2
H. Chen et al. Results in Engineering 24 (2024) 103553

- The hybrid algorithms for data augmentation and transfer learning of hazard, prohibitory, mandatory, and miscellaneous signs. Malaysian
TSD based on integration of the classical methods (feature-based traffic signs can be categorized into guide, Prohibitory, warning signs,
detection and model-based detection) and deep learning methods, and many other traffic-signs as shown in Fig. 3. The above-mentioned
hasn’t yet been fully explored, thus they are presented in this paper three countries have available databases such as German GTSDB data­
with analyzing it effectiveness, advantages and disadvantages. set [23], Chinese TT100K dataset [24], and Malaysian Road Signs in­
- By comparing various algorithms for deep learning in the TSD, this formation [25].
investigation presents the effectiveness of regression, segmentation, By understanding the different categories of traffic signs and their
and hybrid approaches considering metrics such as true positive rate, variations across countries, researchers can develop good detection al­
false positive rate, number of test images, accuracy, processing time, gorithms for recognizing and classifying various types of signs. This
and dataset coverage. comprehensive understanding will also help in the development and
- The previous research has focused on the feature-based detection and evaluation of TSD systems, as well as in the identification of benchmarks
model-based detection methods, such as the use of HOG features or for these systems.
colour- and shape-based detection methods, which encounters many
problems when dealing with complex backgrounds and changing
lighting conditions. Although deep learning has solved many of these Common datasets for traffic sign detection
problems, there are still some issues regarding accuracy and rapid
detection. The development of robust and efficient TSD systems necessitates the
- The paper presents the challenges of current TSD technology, ac­ availability of comprehensive traffic sign datasets. These datasets act as
cording to the latest decade of research advances in the TSD problem, essential tools for the training, validation, and assessment of recognition
and provides future trends for further development of TSD technol­ methods. Typically, the traffic-sign datasets contain a diverse collection
ogy which will help to enrich the TSD research directions. of traffic sign scenarios, portraying various categories of traffic signs,
including guide, interim, regulatory and cautionary indicators.
The remainder of this paper is organized as follows: Section 1 in­ In the recent years, numerous research groups around the world have
troduces the importance of Traffic Sign Detection (TSD) reviews and made concerted efforts to build and expand traffic sign datasets, which
highlights the motivation for this study. Section 2 provides an overview form a cornerstone for advancing the TSD development. Table 1 outlines
of traffic signs and a detailed review of existing TSD datasets. Section 3 a compendium of some of the outstanding datasets that are currently
categorizes traffic sign detection approaches into three primary available to the scientific community, along with their key attributes.
methods: traditional algorithms, deep learning-based methods, and
hybrid approaches. The traditional algorithms focus on early techniques - German Traffic Sign Datasets (GTSDB and GTSRB) [23]: Serving as a
like color segmentation, shape recognition, and pattern matching. The pioneer in this field, the German traffic sign datasets, including
deep learning-based methods leverage advancements in convolutional GTSDB and GTSRB benchmarks, hold a vast collection of 121,000
neural networks (CNNs) and related techniques to significantly improve signs across 43 different categories, captured in 900 images at 640 ×
detection performance. The hybrid approaches combine traditional 480 resolution. This compilation features a variety of mandatory,
methods with deep learning enhancements, such as data augmentation danger, and prohibitory signs, displayed within various traffic
and transfer learning, to further improve accuracy and robustness.
Section 4 discusses the evaluation protocols used to assess the robustness
and performance of various traffic sign detection algorithms. Finally,
Section 5 presents a summary of key findings, offers insights, and pro­
poses directions for future research in the field of TSD.

Overview on the road’s traffic signs and common datasets

The traffic signs on the roads are slightly varied from one country to
another. Thus, there exist several datasets from different countries. The
overview on the traffic signs and the existed datasets are explained here:

Overview on traffic signs on roads

Traffic signs on the roads are purposefully positioned to guide the

drivers with information about upcoming safety concerns, navigational
instructions, restrictions, or text-based details on the directions and
places. Although there’s a noticeable diversity in the design and form of
traffic signs worldwide, the most fundamental classifications of traffic
signs worldwide include the mandatory signs, prohibitory things to be
avoided during driving, hazard and danger situation on the roads, and
text-based indicators on road details. All above-mentioned signs except
the text indicators have commonly consistent shapes such as circles,
triangles, and rectangles, accompanied by universal color codes such as
red, yellow and blue. By contrast, text-centric signs don’t have a
consistent shape and enable to convey instructional contents. In this
section, a comprehensive overview on the different categories of traffic
signs is provided to help researchers to develop a good design and al­
gorithms that accurately and reliably detect and recognize various types
of signs. This understanding will ultimately contribute to safer and more
efficient road navigation. Fig. 3. Various types of traffic sign boards: (a) Common sign boards in Ger­
The German and Chinese based traffic signs can be categorized into many (b) Common sign boards in China, (c) Common sign boards in Malaysia.

3
H. Chen et al. Results in Engineering 24 (2024) 103553

Table 1 drawbacks when dealing with the dataset.

Details of available datasets. The detection of traffic signs is critical for the driver assistance sys­
Dataset Details of the public datasets tems and self-driving vehicles to automatically identify and classify the
traffic signs with real-life scenarios [1,31]. As a vital step in under­
Classes Total Total Image Country
Images Signs Sizes standing and responding to traffic regulations, it has attracted signifi­
cant research interest [32] Examples of challenging scenes from the
GTSDB/GTSRB 43 900 121,000 640*480 German
TSRB 43 5110 11,005 Various Various
CURE-TSD-Real dataset depicting various adverse conditions that
BTSD 62 1506 10,000 640*480 Belgian impact traffic sign detection. The figure shows images of traffic signs
STSD 11 1000 3000 1024*768 Sweden affected by (top row): dirty lens, haze, snow, blur, and rain. The middle
LISA TS 47 7000 47,000 640*480 USA row depicts partial occlusions, shadowing, and physical damage to
CTSD 15 2368 7500 Various China
signs, while the bottom row includes motion blur, low resolution,
TSD IN THE 43 5000 7000 Various Various
WILD overexposure, and underexposure. These conditions represent signifi­
BTS 34 1395 5000 Various Brazil cant challenges for effective detection and recognition in real-world
INDIAN TS 78 1410 8000 Various India scenarios, as shown in Fig. 5 [32,33]. With referring to TSD methodol­
SPANISH TS 47 3300 21,000 Various Spain ogies, one can delineate three main categories: conventional techniques,
MALAYSIA TSR 68 10,000 10,000 Various Malaysia
deep learning-based methods, and integrated or hybrid strategies [34,
35]. Traditional algorithms rely on color segmentation; shape recogni­
scenarios, as shown in Fig. 4. Every traffic sign in these datasets has tion; pattern recognition, and Machine Learning-based detection. Deep
been diligently marked as rectangular regions of interest (ROIs). learning algorithms, especially CNNs, have gained prominence due to
- LISA TS (USA)[26]: A large collection of 47,000 signs in 47 cate­ their capability to learn the hierarchical features and achieve superior
gories, captured in 7000 images at 640 × 480 resolution, providing a performance [2]. Hybrid algorithms attempt to combine the strengths of
rich source of material for research and analysis. traditional or feature-based algorithms with deep learning techniques
- CTSD (China) [24],: A collection of 2368 images in various resolu­ [34,36].
tions, including 15 categories and a total of 7500 signs, epitomizing This paper offers an exhaustive overview of traffic sign detection
traffic signs in China. approaches, highlighting their evolution and the challenges that they
- Malaysian TSR[25]: This dataset includes 10,000 images, comprising are addressed in real-world conditions as follows:
68 categories of traffic signs in Malaysia in different resolutions. This As illustrated in Table 2 Traditional algorithms primarily involve
wide-ranging collection showcases the diversity of traffic signs in the four processes, namely, preprocessing of the live streaming videos,
country. feature extraction, clustering and classification of extracted features,
- Companion datasets such as– TSRB, BTSD (Belgium) [26], STSD and post-processing processes. The conventional algorithms can be
(Sweden) [27], BTS (Brazil) [28], Indian TS (India) [29], and primarily classified into two main classes: feature-based detection and
Spanish TS (Spain) [30] add more depth to the repository, providing model-based detection[38–40]. The feature-based approaches focus on
a versatile platform for researchers to build resilient and adaptable isolating traffic signs based on their distinct colors [37,38], however,
TSD algorithms. model-based techniques rely on matching the predefined templates with
the detected signs [39–42]. Recently, deep learning techniques, espe­
By leveraging such exhaustive datasets, researchers can effectively cially CNNs, are becoming increasingly popular. Their inherent ability to
design, test, and evaluate TSD algorithms. This, in turn, facilitates the independently identify hierarchical attributes positions them as a
advancement of autonomous transportation systems and paves the way prominent solution, which provides exceptional findings in the field of
towards intelligent vehicle technologies. TSD [34,43]. Lastly, hybrid algorithms combine the strengths of tradi­
This section presents available traffic sign datasets and their tional and deep learning approaches [43,44]. For instance, traditional
importance for the development and assessment of traffic sign detection algorithms could be employed for initial traffic sign localization, while
algorithms. Then, various traffic sign detection techniques, their appli­ deep learning is used for classification, or vice versa.
cation with these datasets, and their respective benefits and challenges This paper explores each approach in detail, explaining their tech­
are discussed. niques and the advancements made in the area of TSD over time.
Although hybrid algorithms for TSD are not studied well in the litera­
Traffic sign detection approaches ture, Bayoudh et al., have presented the benefits of utilizing this direc­
tion for road signs detection [18,34]. They leverage the strengths of both
This section will discuss traditional methods, deep learning methods traditional techniques with deep learning algorithms to improve the
and hybrid methods, each of which displays distinctive advantages and traffic sign detection performance. One approach utilizes traditional
algorithms for initial traffic sign localization, such as color segmentation
and shape recognition, while employing deep learning techniques, such

Fig. 4. Example of the traffic scenarios in the German Traffic Signs Detection Fig. 5. Shows shape of sign, Challenging scene examples from the CURE-TSD-
Benchmark(GTSDB). Real dataset: dirty lens, haze, snow, blur, and rain, respectively.

4
H. Chen et al. Results in Engineering 24 (2024) 103553

Table 2
Methods devised for traffic sign detection.
Methods devised for traffic sign detection

Traditional Method Feature-based detection

Model-based detection
Deep Learning Method Regression-based detection methods
Segmentation-based approaches
Hybrid Method Data augmentation techniques
Transfer learning approaches

as CNNs, for the subsequent classification of the detected signs [45,46].

Conversely, deep learning can be used for coarse localization, with
traditional techniques that can be employed for refining the results and
reducing false positives errors [18].By discussing some case studies of
hybrid approaches in traffic sign detection, elaborating on their tech­
niques, strengths, and limitations, one can provide insights into the
potential benefits of combining traditional and deep learning techniques
in the traffic sign detection, as well as identify areas where further
research is needed.

Traditional-based traffic sign detecting algorithms Fig. 6. The Feature-based detection approach[16].

The traditional TSD algorithms focus primarily on techniques such as Feature extraction methodologies, like Histogram-Oriented-
color segmentation, shape recognition, and pattern matching to recog­ Gradients (HOG) [49,56], Scale-Invariant Feature Transform (SIFT)
nize and localize traffic signs in the real-world scenarios. These strate­ [57], or shape descriptors, are then applied to capture the unique
gies aimed to differentiate traffic signs from the background of images characteristics of traffic signs. These features serve as input for classi­
by effectively segmenting and filtering the image based on some distinct fication algorithms, such as Support-Vector-Machines (SVM),
features. In fact, TSD relied mainly on a combination of image pro­ decision-trees, or k-Nearest Neighbors (k-NN), to detect and determine
cessing techniques and some handcrafted features before the widespread the type of traffic signs based on the extracted features [37,58–62].
adoption of deep learning. The selection of the sign boards features
extraction techniques is crucial for both the segmentation and filtering Model-based detection
processes as they should be detected reliably with different image con­ Conversely, the algorithms of model-based detection typically
ditions, including sign’s location, size, and color. Challenges arose due involve the construction of parametric or geometric models representing
to the similarity of color and texture between the boards’ signs and other the appearance and structure of traffic signs, as shown in Fig. 7. [63].
detected objects within a given scene, as well as varying of environ­ These models can be matched against the input image using techniques
mental conditions, represent the major complexity for detecting and like template matching, Hough Transform [64], or Random Sample
recognizing the traffic signs using solely traditional algorithms. Consensus (RANSAC) [65] to identify and locate the traffic signs. The
As has been mentioned earlier, the TSD algorithms are classified into models can be designed to account for variations in scale, rotation, and
two dominant types: feature-based detection and model-based detec­ perspective, enhancing their robustness against diverse viewing condi­
tion. Both approaches have played a crucial role in the initial stages of tion [66]. Nonetheless, developing accurate models capable of handling
TSD, laying the groundwork for the development of more advanced a quite wide-range of variations in the sign appearance and robust
techniques. The details of these strategies, discussing their techniques, matching techniques capable of dealing with occlusions, deformations,
strengths, and limitations in the context of TSD, will be resented as and varying environmental conditions are still a challenging task [67].
follows: Both feature-based and model-based approaches possess their own
strengths and limitations as discussed by [68]. Feature-based algorithms
Feature-based detection are often more robust against varying lighting conditions, occlusions,
The feature-based detection approach emphasizes the extraction and
analysis of distinct features from the image, including color, shape, and
texture, to identify and locate traffic signs, as shown in Fig. 6. This
approach generally involves preprocessing steps to enhance the image,
such as color space conversion, filtering, or histogram equalization
[47–50]. Subsequent to preprocessing, image processing techniques
such as color segmentation, edge detection, or morphological operations
are employed to emphasize traffic sign candidates within the image
[51]. Many studies have used a feature-based detection to extract the
useful parameters from image [39,52–54]. In [55], a preliminary
filtering method was employed first for training the signs color classifier
that will be passed through a linear regression with the equation as
depicted in Eq. (1)

f(x) = (w, x) + b, x = (v1 , v2 , v3 )i (1) Fig. 7. The model-based detection approach[67],(A): Original image showing a
traffic sign.(B) Edge detection applied to extract the contours of the image.(C)
where vi is the intensity value of the RGB-images (i.e., i = 1, 2, 3). The Identification of candidate regions corresponding to potential traffic signs.(D)
governing parameters function are represented by (w, b) ∈ R ×R, where Feature-based model matching to verify candidate traffic sign regions.(E)
w and b are real numbers. The decision rule for this function is deter­ Verified model of the traffic sign overlaid on the original image.(F) Final
mined by the signum function, denoted assgn(f(x)). detection result with precise bounding box around the detected sign.

5
H. Chen et al. Results in Engineering 24 (2024) 103553

and sign degradation, while model-based algorithms can be more ac­ proposal generation for sign candidates or directly utilize the learning of
curate in detecting traffic signs with well-defined shapes and structures. sign pixels in the trained data to regress the corners of bounding boxes
However, both approaches can struggle with complex and cluttered [86]. Several researchers have refined regression-based algorithms to
environments where traffic signs may be partially occluded, degraded, accommodate multi-oriented signs by incorporating regional attention
or share similar characteristics with background objects. These chal­ mechanisms such as the Rotation Region Proposal Network (RRPN),
lenges, coupled with the complexity of real-world traffic sign detection whose detection pipeline is shown in Fig. 11, or Geometric-Aware Sign
scenarios, led to a need of development for more advanced techniques, Detection (GASD), illustrated in Fig. 12. These enhanced methods
such as deep learning-based algorithms, address specific challenges such as rotation or occlusion, but their per­
formance in terms of overall efficiency and accuracy varies widely, as
shown in Table 3, where older and newer models are compared
Deep learning-based traffic sign detection algorithms side-by-side. All the above-mentioned algorithms have been tested with
the GTSRB database, and their results are reported in Table 3.
The development of TSD algorithms has been significantly influ­ Table 3 compares several algorithms, including SSD [88], YOLO[95],
enced by the emergence of deep learning techniques, particularly CNNs,
YOLOv8 [92], EfficientDet, [93] and Faster R-CNN[96], which offer
datasets and various object detection algorithms, as shown in Fig. 8 different approaches—ranging from generating anchors and proposals
[69], Lopez-Montiel assessed the application of deep learning models in
for sign candidates to directly predicting bounding box corners. SSD
embedded systems for detecting traffic signs and concluded that possesses an impressive accuracy rate of 93 %, with a processing time of
employing the Tensor Processing Unit (TPU) can significantly decrease
30–45 ms. On the other hand, YOLO’s strength lies in its ability to detect
training duration[74]. These approaches have brought substantial im­ images in real-time, taking just 20–40 ms but achieving an accuracy of
provements in the TSD field, especially for enhancing the accuracy and
89 %, slightly lower than SSD. The YOLOv8 version improves upon its
resilience elimination amidst the intricated environments. By leveraging predecessors with an accuracy of 93 %, with a processing time of 15–30
large-scale annotated datasets and powerful learning algorithms, Deep
ms, showing both enhanced speed and reliability. EfficientDet is note­
learning models inherently grasp hierarchical attributes and consis­ worthy for its balanced accuracy-to-speed ratio, offering an accuracy
tently deliver cutting-edge outcomes in the realm of traffic sign detec­
rate of 94 % and a processing time of 25–50 ms, thereby demonstrating
tion [70–72]. its capacity to operate efficiently in constrained computational envi­
In this section, three primary deep learning-based TSD methodolo­
ronments. DETR represents another significant development, offering an
gies are presented, namely, regression-based, segmentation-based, and
end-to-end detection approach that achieves a 95 % accuracy rate,
hybrid approaches as shown in Fig. 8.
though at a slightly higher computational cost of 60–120 ms due to its
The deep learning-based methodologies for TSD typically involve a
transformer-based architecture[90]. For the tasks that require high ac­
combination of classification, regression, and template matching. This
curacy, the Faster R-CNN shines with an impressive true positive rate of
workflow is visually represented in Fig. 8, which shows the overall
97 %, albeit with a higher computational time of 60–150 ms. Other al­
procedure of a deep learning algorithm applied to TSD. The figure
gorithms such as RRPN [91], and GASD [87] have enhanced the algo­
demonstrates how the CNN extracts features from an input image, which
rithmic landscape by addressing the niche requirements, such as
are subsequently used for both classification (assigning a shape label)
detecting rotated objects and managing occlusions, respectively. All
and regression (determining the 2D pose). These steps collectively
above-mentioned algorithms have been validated using GTSRB data­
enable accurate detection and localization of traffic signs, followed by a
base, confirming their relevance and applicability in real-world
template transformation step to refine boundary detection.
scenarios.
Yang et al., presented a TSD algorithm that uses the Attention-
Regression based detection algorithms
Network (AN) and Faster-RCNN to recognize the ROIs and cluster the
The regression algorithms for detecting the traffic signs, such as You
traffic signs [3]. The final region proposals are determined using
Only Look Once (YOLO) and Single Shot Detection (SSD) [44,73–80]. A
Fine-Region-Proposal-Network (FRPN). This approach aims to enhance
comparison between SSD and YOLO is illustrated in Fig. 9 [81,82].
the precision and efficacy of TSD in real-life scenarios. Ma et al., pro­
Additionally, YOLOv8 has recently been introduced, further improving
posed a Rotation-Region-Proposal-Networks that employs a
upon the real-time performance characteristics of its predecessors while region-proposal-based approach to create sloping proposals for the text
enhancing accuracy and robustness in complex environments[88].
orientated based angle (Ma et al., 2018). This approach improves the
Another significant development is EfficientDet, a detection model that fitting of text regions and facilitates rectification, leading to enhanced
employs a compound scaling method to balance accuracy and efficiency.
text reading performance. However, the main limitation of
This approach allows the model to adapt to different resource environ­ regression-based approaches is their inability to effectively localize signs
ments, making it particularly valuable in situations where computa­
with a large aspect ratio due to the convolution network’s constrained
tional power may be limited but performance must remain optimal[89]. receptive field.
DETR (DEtection TRansformer), which uses a transformer-based archi­
tecture, presents a significant innovation by eliminating the need for Segmentation-based approaches
anchor boxes, providing an end-to-end detection framework suitable for
Segmentation-based detection algorithms encompass two general
complex scenarios [83].Another algorithm for Regression-based detec­ categories, namely, instance and semantic segmentations. A summary of
tion is Faster Region CNN (Faster R-CNN) which has the architecture as
various TSD systems testing using segmentation-based detection algo­
shown in Fig. 10 and consider traffic signs as specific objects to be rithms in GTSRB dataset, including their detection features, pros, merits,
detected within an image [84,85]. These newer models, YOLOv8, Effi­
true positive rate, false positive rate, test images numbers, cumulative
cientDet, and DETR, alongside traditional ones, employ either anchor or accuracy, duration, and associated dataset, is presented in Table 4.
Semantic segmentation techniques, such as Pixel Aggregation
Network (PAN) and Mask Guided Pixel Aggregation Network (MGPAN),
classify each pixel in an image as a sign or non-sign. These algorithms
rely on the geometric properties of signs to derive the bounding box, yet
they may encounter difficulties when processing dense signs with
identical orientation [18,38,73,100,101]. J. Wu et al., have combined
SSD with Receptive Field Module (RFM), and Path Aggregation Network
Fig. 8. The overall procedure of the deep learning algorithm [69]. (PAN) [79]. PAN integrates multiscale features, enhancing the

6
H. Chen et al. Results in Engineering 24 (2024) 103553

Fig. 9. Example of A comparison between SSD and YOLO structure[94].

Fig. 10. Example of Architecture of Faster R-CNN[23].

of the traffic signs, resulting in improved system performance.

Soilán et al., presented a novel algorithm for automated identifica­
tion of the abnormal traffic signs in 3D-point clouds, such as the signs
located in vertical directions, using a LYNX Mobile-Mapper system
[102]. Such an approach incorporates machine learning techniques to
recognize the traffic sign semantics from 2D images generated by
reprojection of the 3D positions. By combining LiDAR point clouds and
image analysis, this algorithm provides a powerful tool for enhancing
traffic sign detection and inventory management. Kamal et al., consid­
ered solving of the TSD through the paradigm of image segmentation
and a new type of SegUNet, both are integrated with U-Net and CNN
Fig. 11. Example of Rotation-based text detecting pipeline [99].
architectures to recognize the traffic signs within the live-streaming
video sequences [38]. They train the SegU-Net using CURE-TSD data­
recognition capability of the features and improving the detection ac­ base with around 1.7 millions challenge-free traffic-scene images, using
curacy. The spatial pyramid pooling module pools shallow features and a Tversky-loss-function with an L1 term to detect small signs, achieving
supplements the feature system, further improving detection perfor­ 73 % precision and 61 % recall. An approach that enhances the images
mance. Y. Zhu et al. proposed a dual-phases based algorithms for the adaptively through Gamma correction and contrast-limited histogram
traffic sign text detection that utilizes a fully-CNN and Fast-NN to equalization have been introduced by [99], which uses MobileNet-V2
enhance the efficiency and accuracy of text recognition [96]. This al­ and DeepLab-V3+ algorithms for segmenting the traffic signs and
gorithm segments the candidate of traffic sign areas with ROI, followed marking them in multiple scenes.
by text detection within the extracted regions. The dual-phase detection Path Aggregation Network (PANet) has been proposed to enhance
methodology decreases the searched area and removes the text outside the flow of information in the framework of instance segmentation

7
H. Chen et al. Results in Engineering 24 (2024) 103553

Fig. 12. Example of Architecture of the GAPretrain [87].

Table 3
TSD systems using Regression-based detection algorithms.
Ref Algorithm Pros Cons Application True False No. of Accuracy Duration Dataset
Scenario Positive- Positive- Test
Rate Rate Images

Wei Liu SSD Fast/scalable/ Higher false Suitable for 0.94 0.08 500 0.93 30–45 GTSRB
et al., efficient positive rate scenarios (ms)
[88] requiring speed
Joseph YOLO Extremely fast/ Lower accuracy in Real-time 0.9 0.12 500 0.89 20–40 GTSRB
Redmon real-time complex scenes applications (ms)
et al.,
[89]
KH Shih Faster R-CNN High accuracy/ Longer processing Applications 0.97 0.05 500 0.96 60–150 GTSRB
et al., robust time requiring high (ms)
[90] precision
Van-Dung Rotation Detects rotated Longer duration Scenarios with 0.95 0.07 500 0.94 80–200 GTSRB
Hoang Region objects variable sign (ms)
et al., Proposal orientation
[91] Network
Linyan Geometric- Handles occlusions, Processing time Detecting 0.96 0.06 500 0.95 70–180 GTSRB
Huang Aware Sign context-aware needs optimization occluded or (ms)
et al. Detection partially visible
[87], signs
[92] YOLOv8 ​ ​ ​ ​ ​ ​ ​ ​ ​
​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​
​ High speed Complex to train Real-time 0.92 0.06 500 0.93 15–30 GTSRB ​
and accuracy applications with (ms)
challenging
conditions
[93] EfficientDet Balanced accuracy/ Relatively complex Scenarios needing 0.94 0.05 500 0.94 25–50 GTSRB
computational architecture an accuracy-to- (ms)
efficiency speed balance
[83] DETR End-to-end Requires more Large-scale 0.95 0.04 500 0.95 60–120 GTSRB
detection, no computational detection tasks (ms)
anchor boxes resources with complex
needed backgrounds

based neural network [103]. This can be achieved through enhancement noteworthy efficiencies within the segmentation-based traffic sign
of the feature’s hierarchy and high precise data location for the lower detection paradigm on the GTSRB dataset. With its proficient
layers of the data via augmented bottom-up path. Subsequently, an contextual-based mechanism, the Mask-Guided-Pixel-Aggregation-
adaptive-based pooling features that connect the features’ grids to other Network (MGPAN)[109] achieves a superior accuracy of 96 % within
hierarchical features levels, is established. Instance segmentation tech­ 85ms processing time. The Pixel-Aggregation-Network (PAN) and the
niques, exemplified by Pixelink and Mask-RCNN, segment sign instances Path-Aggregation-Network (PANet) exhibit commendable accuracies of
by treating sign mask candidates as connected components [104,105]. 94 % [79]. However, PANet’s processing duration is marginally
Contemporary studies have explored the application of the extended into 90ms compared to the origin PAN’s with a processing time
Fourier-Contour-Embedding (FCE) algorithm for detecting non-standard of 80ms. The Fully-Convolutional-Networks (FCN)[102] highlight its
shape signs, capturing sign features without necessitating of additional rapidity with processing time of 40ms, although it compromises slightly
data during the training process [27,95], while it is not able to adapt to on accuracy, achieving 91 %. With a processing time of 50ms, SegNet
the different sign shapes, due to that the embedding process is fixed [38] attains to reach an accuracy of 90 %, highlighting its streamlined
during the training phase and cannot be modified to accommodate new architecture. In contrast, the Fourier-Contour-Embedding (FCE) algo­
shapes. rithm [95] emphasizes the rotation-invariant features and achieves a 93
Several algorithms in Table 4 have demonstrated diverse yet % accuracy within 75ms processing time. The Mask-R-CNN [97], which

8
H. Chen et al. Results in Engineering 24 (2024) 103553

is popular for its instance-aware precision, achieves an impressive of 95

Dataset
% accuracy, albeit within a longer processing interval of 200ms. Models

GTSRB

GTSRB

GTSRB

GTSRB

GTSRB

GTSRB

GTSRB

GTSRB

GTSRB
like DeepLab [99] has a boundary sensitivity and performs closely to
FCE and PAN with a 93 % accuracy and 60ms processing time. U-Net
[98], which is praised for its parsimonious architecture, achieves an
Duration

200ms
80ms

85ms

75ms

40ms

50ms

90ms

45ms

60ms
accuracy of 92 % and 45ms processing time.
To summarize, some algorithms prioritize detection precision, like
MGPAN and Mask R-CNN, while others, like FCN and U-Net, prioritize
Accuracy

the rapid processing. Therefore, when choosing these algorithms for a

94 %

96 %

93 %

91 %

90 %

95 %

94 %

92 %

93 %
specific application, it is important to carefully evaluate the trade-offs
between accuracy and speed.
No. of Test

Fused deep learning-based detection algorithms

Images

The fused deep learning-based detection algorithms amalgamate

500

500

500

500

500

500

500

500

500
elements from both regression-based and segmentation-based ap­
proaches, leveraging the strengths of each methodology, As shown in
Positive-Rate

Table 5. In this algorithm, contour modeling can be integrated with

regression-based sign detection to identify the arbitrary-shaped signs
10 %
False

7%

6%

8%

9%

5%

7%

8%

7%
[106–109]. Polewski et al., presented Contour Regression-based Detec­
tion (CRD), in which, the framework is designed to segment the in­
stances of the object’s class using a multi-layer active of contour-based
evolution in the form of semantic segmented maps [110]. This frame­
Positive-

work leverages on the fully-CN and frames contour-evolution with

94 %

96 %

93 %

91 %

90 %

95 %

94 %

92 %

93 %
True

Rate

encompassing a data fit-term, to determine the shape model, and object

overlapping. The efficient-solution-neighborhood-operators (ESNO) are
Context-heavy areas, such as

Situations requiring precise

Scenarios with diverse sign

proposed for optimizing the object recognition using metaheuristic

High-traffic areas, complex
Lightweight applications,
Areas with dense traffic

techniques like simulated-annealing. Additionally, the utilization of

Quick deployment, less

Boundary detection in
Varied environments,
adaptable to changes
Application Scenario

Fully CN in combination with the Feature-Pyramid-Network (FPN)

mobile deployment

diverse conditions

makes it possible to combine the features of both low and high levels in
computational
urban scenes

the image. This enables the framework to recognize signs of various

orientation

backdrops
detection

sizes. [111–113]. In order to guarantee the accuracy of traffic sign

detection, the S3B-Net algorithm was presented by [114], which is
signs

working based on Segmentation-Proposal Networks (SPNets) and

instance segmentation algorithms for segmenting the object beyond the
High complexity, more parameters
Less effective on non-contour signs

Potentially high false positives on

Can be computationally intensive

May not be as robust in variable

bounding box. The algorithm designs a sub-network to aid instance

May struggle with non-standard
May require tuning for scale

segmentation based on object detection. The sub-network predicts firstly

May oversimplify complex

the embedded pixels within the network, then it calculates the proba­
cluttered background
computational power

bility of each pixel belonging to an instance using a Gaussian function.

Requires substantial

Finally, the output is combined with instance segmentation of the object

for accurate instance-mask. Qian et al., leveraged the accurate region
conditions

identification from semantic segmentation and the precise localization

scenarios
variance

needed

of bounding box regression to improve overall detection performance

signs
Cons

[115]. The amalgamation of these methods has the potential to bolster

the performance of TSD, enhancing both precision and operational ef­
Examples of TSD systems using Segmentation-based detection algorithms.

ficiency. However, the actual performance of the fusion semantic seg­

Precise, instance-aware
Dense pixel predictions

Fast, fewer parameters

Shape-based, rotation-
Contextual, improved

Accurate, boundary-

mentation and bounding box regression (FSS-BBR) algorithm will

Good performance,
Hierarchicafeature

Enhanced feature

depend on the specific implementation, data pre-processing, and

training procedures [116]. The work introduced by [117] indicated that
propagation
lightweight
invariant

the main idea behind multi-task learning is to enable the network

accuracy

sensitive
fusion

learning from the shared features and representations that are useful for
Pros

multiple tasks, leading to improve overall performance. However, the

actual performance of the MTLN will depend on the specific imple­
Fourier Contour Embedding
Pixel Aggregation Network

Path Aggregation Network

mentation, data pre-processing, and training procedures. A research

released by [118] has highlighted that Spatial Attention-based Detection
Aggregation Network

Fully Convolutional

(SAD) can be incorporated with spatial attention mechanisms, allowing

Mask Guided Pixel

Networks (FCN)
(FCE) algorithm

the network to concentrate on the relevant regions in the image. This

Mask R-CNN

leads to a precise TSD by reducing the false positives and minimizing the
Algorithm

DeepLab
(PANet)

impact of background clutter. In fact, the actual performance of the SAD

SegNet
(PAN)

U-Net

algorithm will depend on the specific implementation, data

pre-processing, and training procedures. Table 5 shows a comparison
between multiple hybrid algorithms in GTSRB dataset.
U Kamal et al. [38],
Yingying Zhu et al.
Jianjun Wu et al.,

Shuen Zhao et al.

Jianjun Wu et al.
Xijun Qian et al.

Xiaomei Li et al.

Table 5 evaluates the hybrid deep learning-based detection algo­

Yiqin Zhu et al.,

O Ronneberger
et al. [98],

rithms using the GTSRB dataset by combining regression and segmen­

tation methodologies to achieve the optimal detection outcomes.
[94],

[96],

[97],

[79],

[99],
Table 4

[79]

[95]

CRD demonstrates its ability to handle arbitrary-shaped signs, with a

Ref

93.5 % accuracy and a processing duration of 150ms. The SPNets

9
H. Chen et al. Results in Engineering 24 (2024) 103553

Table 5
Examples of TSD systems using Hybrid detection algorithms.
Ref Algorithm Pros Cons Application Scenario True False No. of Accuracy Duration Dataset
Positive- Positive- Test
Rate Rate Images

Polewski Contour Handles May have Areas with diverse 94 % 6% 500 93.5 % 150ms GTSRB
et al. Regression-based arbitrary shapes difficulties with sign shapes and
[110], Detection (CRD) highly cluttered placements
backgrounds
Jinbao Segmentation- Unified May require Environments 93 % 7% 500 92.5 % 170ms GTSRB
Zhang Proposal segmentation- complex needing integrated
et al., Networks proposal configuration detection and
[114] (SPNets) segmentation
Jinhui Hou Multi-Task Joint learning, Be challenging to Scenarios 95 % 5% 500 94.5 % 140ms GTSRB
et al. Learning shared features balance multiple demanding
[117], Networks tasks simultaneous feature
(MTLN) extraction and
detection
Fenglei Spatial Attention- Focused regions, High computational Scenarios 96 % 4% 500 95 % 130ms GTSRB
Ren et al. based Detection improved load for real-time demanding
[118], (SAD) efficiency processing simultaneous feature
extraction and
detection

proposes a unified segmentation-proposal mechanism that attains an Hybrid algorithms for TSD
accuracy of 92.5 % with a time processing of 170ms. Multi-Task
Learning Networks (MTLN) achieves a commendable of 94.5 % accu­ The development of deep learning algorithms has shown a great
racy within a duration of 140ms, emphasizing its potential in utilizing impact on the advancement of traffic sign detection algorithms, leading
well the shared features. Significantly, the SAD algorithm records the to notable improvements in performance in comparison with conven­
highest accuracy of 95 % in a timeframe of 130ms by utilizing spatial tional approaches. In this review, a hybrid algorithm for traffic sign
attention modality, highlighting its effectiveness in region-focused detection (TSD) refers to the integration of data augmentation and
detection. The results indicate that the combination of regression and transfer learning. Data augmentation is used to artificially enhance the
segmentation approaches can increase the accuracy and efficiency of diversity of the training dataset by creating variations of existing data,
traffic sign detection. which helps the model generalize better to new situations. Transfer
Recent advances in two-stage detection frameworks have proven learning, on the other hand, enables the use of pre-trained networks to
effective in enhancing the accuracy of traffic sign detection. [119] adapt to a new but related task, reducing computational resources and
proposed a method for traffic light detection and recognition using a training time while improving model performance. Together, these
two-stage framework, which involves individual signal bulb identifica­ methods form a powerful hybrid approach to TSD, providing enhanced
tion to improve precision during adverse conditions such as glare or robustness and accuracy.
occlusions. This method emphasizes the detailed recognition of indi­
vidual components, leading to increased detection reliability in chal­ Hybrid methods that integrate classical preprocessing techniques
lenging environments. [120] introduced a two-stage fusion neural One form of hybrid approach involves the integration of classical
network approach that aims to integrate features from multiple stages to machine learning methods—such as color segmentation or edge detec­
better detect and classify traffic signs. This fusion of data across stages tion—with deep learning models to leverage the strengths of each.
allows for enhanced robustness against varying lighting conditions and Classical techniques, such as color segmentation or edge detection, are
occlusions in real-world applications. efficient for preprocessing tasks that localize candidate regions of in­
Deep learning-based traffic sign detection algorithms have demon­ terest. These candidate regions are then processed by deep learning
strated significant advancements over traditional techniques, yielding to models (e.g., Convolutional Neural Networks (CNNs)) for refined clas­
improve the TSD tasks. These methodologies can address a wide spec­ sification and localization.
trum of traffic sign variations and robustly detect signs in intricate and In TSD, color-based segmentation can be used as a preprocessing step
cluttered environments, which can effectively overcome the limitations to isolate regions in an image that are likely to contain traffic signs. For
of traditional feature-based and model-based algorithms. instance, many traffic signs are characterized by specific colors (e.g., red
In this section, we compare the different approaches to Traffic Sign for warning signs). By using classical color thresholding [4], the algo­
Detection (TSD) outlined in Tables 3, 4, and 5. Traditional methods (e.g., rithm can efficiently identify regions of interest that are then passed to a
feature-based detection and model-based detection) are computation­ CNN for further classification. This not only reduces computational
ally less demanding and can be implemented with lower resource re­ overhead but also enhances detection accuracy by focusing the deep
quirements; however, they often lack robustness in complex scenarios learning model on the most relevant areas of the image. By integrating
involving variations in illumination and occlusion. Deep learning-based classical preprocessing with deep learning, the hybrid approach com­
methods, such as regression-based and segmentation-based algorithms, bines the efficiency of classical methods with the accuracy of deep
leverage advanced neural networks, significantly enhancing accuracy learning, making it particularly useful in scenarios with limited
and adaptability. Their limitations mainly stem from their higher computational resources or in environments with complex backgrounds.
computational requirements and the need for extensive labeled datasets.
Hybrid methods combine elements of both traditional and deep learning Hybrid methods that integrate classical preprocessing techniques
approaches to achieve a balance between accuracy and computational The hybrid algorithms discussed in this sub-section focus on
efficiency. They are often designed to incorporate the advantages of combining data augmentation and transfer learning—two well-
both approaches, such as leveraging data augmentation and transfer established deep learning strategies that enhance model performance
learning, but their complexity can sometimes limit their applicability, in TSD. These deep learning-based enhancements are particularly
especially in real-time or resource-constrained environments. valuable when the availability of training data is limited or when

10
H. Chen et al. Results in Engineering 24 (2024) 103553

training from scratch would be computationally infeasible. [129] proposed a versatile technique based on the Gaussian mixture
model, coupled with automatic partitioning and transfer learning, to
Data Augmentation address the challenge of traffic sign degradation. This method surpasses
Data augmentation is a technique used to expand artificially the the performance of numerous leading-edge and contemporary algo­
training dataset to cover most of TSD possibilities by applying many rithms, While the technique may be able to handle signs with a certain
types of transformations to the original images, such as re-scaling, degree of degradation, it may struggle to accurately detect or recognize
rotation, flipping, and brightness adjustments, as shown in Fig. 13 signs that are heavily degraded or damaged, Additionally, the described
[121].These transformations can effectively lead to training data di­ technique may require significant computational resources in order to
versity, facilitating the development of models that lead to novel situ­ effectively detect and recognize traffic signs in real-world scenarios. This
ations and mitigating the risk of overfitting [122–126]. could be a disadvantage in situations where resources are limited or the
In the realm of TSD, data augmentation proves particularly benefi­ real-time experiments are critical, such as in autonomous driving ap­
cial for addressing challenges associated with variations in viewpoint, plications. In the study accomplished by [46], a novel approach based on
changes in illumination, and occlusions. By incorporating the data transfer learning is presented, which involves a pre-training model
augmentation, traffic sign detection models can be trained to recognize based on the German traffic-sign dataset and subsequently fine-tuning
signs under an extensive array of conditions, leading to improvements in model using a Pakistani dataset comprising of 359 images collected
both performance and robustness. from various locations in Pakistan.
According to [127], the results show that Mosaic with Mixup algo­ The amalgamation of data augmentation and transfer learning
rithm can bring growth points on a strong augmentation is more helpful. techniques constitutes a hybrid approach that leverages the strengths of
However, this algorithm has a limited effectiveness on the complex both algorithms in traffic sign detection. By employing data augmen­
backgrounds. The study conducted by [128] indicated that R-CNN can tation, the model is exposed to a more diverse set of traffic sign sce­
effectively capture the features with multiple scales in the pyramids and narios, enhancing its ability to detect the previously unseen conditions.
amalgamates them using dot-product and SoftMax operations. This Concurrently, transfer learning harnesses the knowledge gained from
process emphasizes the features of traffic signs, leading to enhanced pre-training on a large-scale dataset, resulting in accelerating the
detection accuracy. As discussed by [35], they extend their data training and improving the performance. This hybrid algorithm has
augmentation model by introducing a pipeline rendering to create a proven to be a powerful approach in the development of TSD systems,
suitable traffic-sign boards on images’ sequences with more realistic leading to advancements in accurate, robust, and efficient models [109,
backgrounds and artifacts. Nonetheless, the cartoon generation process 123,130–134].
will remain inefficient and poorly generated if the low quality of images The study by [62,135], introduced a new data-driven system that
sequences is existed. As a result, pose determination and image’s leverages the strengths of both data augmentation and transfer learning
background segments may become poor, and the entire generation techniques to identify every classification of traffic signs, has been
process may fail to produce satisfactory results. introduced. This encompasses both symbol-driven and text-centric signs
within video sequences recorded by a camera affixed to a vehicle. The
Transfer learning amalgamation of these algorithms constitutes a hybrid approach that
Transfer learning techniques capitalize on pre-trained models or has shown promising results in traffic sign detection, as shown in Fig. 14.
features learned from one task and adapt them into a new, related task. [34] presented a fusion of 2D-3D CNN models, drawing upon the
This approach substantially required training data and computational transfer learning framework, to enhance performance on established
power for the following task, as the model has learned basic charac­ real-world datasets. In their recent study [85], presented a novel
teristics from previous engagement. In the context of traffic sign methodology for constructing a comprehensive Mexican traffic sign data
detection, transfer learning is commonly applied by initializing the set and employs an adapted residual neural network tailored for traffic
model with weights from a pre-trained network, such as ImageNet, and sign categorization. To enhance the efficacy of the neural network
subsequently fine-tuning the model on traffic sign data. This process model, Castruita Rodríguez leverages data augmentation and transfer
enables the model to benefit from the knowledge acquired during the learning techniques, which constitute a hybrid approach that combines
pre-training phase, leading to faster convergence and superior perfor­ the strengths of both algorithms in traffic sign detection. The results of
mance in many cases.

Fig. 13. Image data-augmentation methods. The coloured lines denote the
relationship between a given data-augmentation technique and its corre­
sponding meta-learning approach [121]. Fig. 14. Pipeline of synthesizing traffic signs. [62,135].

11
H. Chen et al. Results in Engineering 24 (2024) 103553

this study underscore the efficacy of the suggested method in accurately In this context, TP represents instances in which the model correctly
recognizing and sorting multiple Mexican road signs. identified the positive class. FP denotes instances where the model
The amalgamation of data augmentation and transfer learning in a incorrectly classified a non-positive instance as positive, while FN refers
hybrid algorithm allows for an effective combination of complementary to cases where the model failed to identify a positive instance correctly.
strengths. Data augmentation increases the diversity of the training data The model’s effectiveness was assessed using the F1 score.
by applying transformations such as rotation, scaling, and brightness
adjustments. This variety helps reduce the risk of overfitting by exposing Intersection over Union (IoU)
the model to a broader range of input scenarios. Meanwhile, transfer
learning utilizes pre-trained models—such as those based on Image­ The Intersection over Union (IoU) is a widely used to gauges the
Net—allowing the model to start from a well-informed point with fea­ overlapping metric that quantifies the degree of interference between
tures already learned from large datasets, thereby accelerating the predicted bounding boxes generated by object detection models and
convergence, and enhancing performance on the target task. the actual annotated bounding boxes which are provided in the original
The hybrid algorithms presented in this section reflect the growing annotated dataset [141,142]. To evaluate detection accuracy, the IoU
trend of integrating diverse techniques to address the specific challenges calculates the ratio of the overlapping region, the intersection between
of TSD, such as limited data availability, complex environments, and the predicted (pred) and ground-truth (gt) bounding boxes over their
computational constraints. The combination of classical methods with union is calculated in Eq. (5) [143,144]:
deep learning models offers a balanced approach that utilizes the
computational efficiency of classical techniques alongside the accuracy IoU = (pred ∩ gt)/(pred ∪ gt) (5)
and robustness of deep learning. Similarly, the combination of data In a related application, researchers have used Intersection over
augmentation and transfer learning within the deep learning paradigm Union (IoU) as an evaluation metric for object detection. In[151], an
forms an effective hybrid approach to tackle overfitting, enhance attentive semi-anchoring guided network uses a customized IoU-based
generalization, and reduce computational demands. mechanism to refine the localization of detected traffic signs. This
modification effectively balances the attention given to different types of
Evaluation of traffic sign detection bounding boxes, thereby improving detection accuracy, especially in
challenging environments where signs may be partially occluded or vary
Deep learning has transformed traffic sign detection using hybrid significantly in shape and size,as shown in Fig. 15. Another approach
models that integrate data augmentation and transfer learning, thereby highlights the use of IoU within the YOLOv8 architecture for multiscale
boosting accuracy. This breakthrough is instrumental in establishing TSD scenarios. This implementation ensures improved performance by
robust and efficient detection systems. Following this, we will explore integrating a tailored IoU mechanism that can accommodate the varying
the essential evaluation techniques for these models, their training scales of traffic signs and their diverse appearances, which are common
datasets, and the choice of evaluation metrics. challenges in TSD datasets [152],as shown in Fig. 16.
Two critical aspects of developing an effective traffic sign detection Recent advancements in traffic sign detection, such as the improved
(TSD) model are selecting suitable evaluation metrics and ensuring sparse R-CNN by [145] and the TSD-YOLO model by [92], have focused
robust testing under real-world conditions. In this section, we will focus on enhancing accuracy for small objects in real-time scenarios. Both
on the evaluation protocols used to measure the success of TSD algo­ methods incorporate advanced attention mechanisms and loss func­
rithms, including metrics such as accuracy, precision, recall, F1 score, tions, yielding significant performance improvements, particularly in
and processing time. These metrics are essential for understanding and complex driving environments.
comparing the effectiveness, efficiency, and robustness of different TSD
approaches.
Mean Average Precision (mAP)
The evaluation of TSD algorithms is crucial to measure their per­
formance, identify areas for improvement, and compare different
Mean Average Precision (mAP) is another popular evaluation mea­
methodologies. In this section, we will discuss the commonly evaluation
surement for object identification endeavors. It is the mean of the
metrics and techniques employed in the assessment of TSD systems.
average precision values for every class of traffic signs, considering both
Suitable diagrams and tables are incorporated to provide a visual
the precision and recall at different detection thresholds. mAP provides
depiction of the examination process.
a single performance measure that can be used to compare different TSD
algorithms. Dewi et al., examined how the object detection algorithms,
Precision, Recall, and F1-Score Yolo V4 and Yolo V4-tiny, perform when combined with Spatial Pyra­
mid Pooling (SPP). For TSD and recognition using Mean Average Pre­
Precision, recall, and F1-score are widely useful metrics for evalu­ cision (mAP) as key evaluation metric [111]. The authors analyzed the
ating classification and detection algorithms [111]. Precision is defined impact of the SPP principle on improving the feature extraction capa­
as the proportion of traffic signs that are detected accurately. The recall bilities and object feature learning effectiveness of the Yolo V4 and Yolo
shows a metric that defines the proportion of actual positives that were V4-tiny backbone networks. By incorporating the SPP into the Yolo V4_1
correctly identified relative to the total positives in the dataset. The model, the algorithm achieved a mAP of 99.32 % which progressive
F1-score combines both precision and recall in a balanced manner. methods. techniques and demonstrates the effectiveness of the mAP
providing a balanced measure of both metrics. The precision (P), and metric in assessing the accuracy and efficiency of TSD algorithms.
recall (R) can be calculated in the subsequent manner [136,137], F1 The Average Precision (AP) is a metric that quantifies the perfor­
[138–140]: mance of an object detection algorithm for a specific class. The measure
TP is calculated by evaluating the area underneath the precision-recall
Precision = (2) curve, as denoted by Eq. 6:
TP + FP
∫1 ∑
AP
TP AP = p(r)dr, mAP = (6)
Recall = (3) NC
TP + FN 0

2⋅Precision⋅Recall where p(r) denotes the precision at a given recall r, NC presents the class
F1 = (4) category, and N denotes the number of images (C). AP is computed
Precision + Recall
separately for every class, resulting in many AP values as there are

12
H. Chen et al. Results in Engineering 24 (2024) 103553

Fig. 15. Example of predicted bounding boxes with and without the feature alignment module.

Fig. 16. Example of visualization of traffic sign detection.

classes in the dataset. The Mean Average Precision (mAP) represents the predicting the majority class.
mean of these AP values across all classes. A data augmentation algorithm based on traffic sign logos is pro­
posed by [125] to confront the disparity in sample counts with traffic
Classification accuracy sign classification. The testing dataset, derived from the TT100k dataset,
comprises 143 categories and 19,546 images. The training dataset is
Classification accuracy refers to the proportion of traffic signs in the generated using the data augmentation algorithm, resulting in 143
dataset that were accurately classified. This metric is commonly used to categories and 715,000 images. Three classifiers, namely, Darknet19,
assess the efficacy of algorithms of traffic sign recognition algorithms. ResNet50, and VGG-16, were trained and tested, achieving classification
This metric is commonly used for traffic sign recognition tasks, which accuracies of 91.3 %, 91.2 %, and 86.4 %, respectively, as illustrated in
involve assigning the correct label to a detected traffic sign. However, it Fig. 17.
is essential to consider the dataset’s class imbalance when interpreting
classification accuracy, as high accuracy could be achieved by merely

Fig. 17. Example of The detection results of: (a) [24], (b)YOLOv3(512) [146] and (c)[131] (512,1024).

13
H. Chen et al. Results in Engineering 24 (2024) 103553

Receiver operating characteristic (ROC) curve and area under the curve they experience a slight decline in accuracy when faced with complex
(AUC) scenarios. Conversely, algorithms like Faster R-CNN exhibit a notable
advantage in terms of accuracy. Nonetheless, they have a protracted
The Receiver Operating Characteristic (ROC) curve graphically processing time and should only be employed in high-accuracy
represents the true positive rate (often termed sensitivity or recall) in demanding situations. Segmentation-based techniques, like Pixel Ag­
relation to the false positive rate (1-specificity), across different detec­ gregation Network (PAN) and Mask R-CNN, offer benefits in detail
tion thresholds. The Area Under the Curve (AUC) is the area under the processing and boundary recognition. However, they may encounter
ROC curve, as shown in Fig. 18., which provides a single value sum­ computational difficulties while handling large-scale data. These ap­
marizing the performance of a TSD algorithm across all detection proaches are highly suited to complex traffic environments that neces­
thresholds. A higher AUC value indicates better algorithm performance. sitate precise boundary identification. Hybrid Algorithms: standing out
[147] pointed that the Receiver Operating Characteristic (ROC) analysis by leveraging a combination of deep learning enhancements such as
has frequently been used as a valuable evaluation technique to assess the data augmentation and transfer learning, as well as integrating classical
detection performance of object detection algorithms. The ROC curve, preprocessing techniques like color segmentation or edge detection. This
depicted in Fig 18. (a), is plotted as a function of true positive rate (TPR) combination offers multiple advantages for different environments:
versus false positive rate (FPR) for a range of detection thresholds. The
Area Under the Curve (AUC) is another widely used evaluation metric. It - Addressing Environmental Complexity: The integration of classical
quantifies the area beneath the ROC curve. methods (e.g., color segmentation) helps in efficiently localizing
The utilization of AUC offers several advantages over the ROC candidate regions, which can be highly useful for scenarios with
curves. For instance, there may be cases where two distinct detectors, A specific color patterns, such as traffic signs. This allows deep learning
and B, produce different 2-D ROC curves, yet both possess the same AUC models to focus computational resources only on relevant areas,
value, as illustrated in Fig. 18(b). In such scenarios, despite having which significantly enhances performance in challenging environ­
distinct ROC curves, both detectors A and B demonstrate equivalent ments with complex backgrounds and varying lighting conditions.
detection performance. The AUC metric provides a singular perfor­ Data augmentation helps mitigate the issue of insufficient training
mance measure that facilitates a more straightforward comparison of data, which is common in scenarios with rare traffic signs or
different object detection algorithms, including those employed in TSD occluded views. By artificially expanding the diversity of training
systems. datasets, data augmentation increases the ability of models to
The evaluation of TSD algorithms is essential for the ongoing generalize effectively across different environments, making them
improvement and comparison of different methodologies. By employing robust to variations like different viewpoints, illumination changes,
rigorous evaluation metrics and techniques, researchers can effectively and occlusions.
assess the performance of their algorithms, identify areas for enhance­ - Improving Detection Robustness: Transfer learning allows models to
ment, and play a role in the progression of robust and efficient TSD leverage the feature extraction capabilities of pre-trained networks,
systems. such as those trained on ImageNet. This enables hybrid algorithms to
Through the use of precise evaluation metrics and techniques, re­ adapt quickly to new but related tasks, significantly reducing the
searchers can evaluate their algorithms effectively, identify specific computational burden and speeding up training. This is particularly
areas for improvement, and contribute towards building strong and advantageous in resource-limited environments, where training from
efficient TSD systems. Moving forward, we will now progress to the scratch may not be feasible.
concluding section of this paper where we will explore future trends. In - Balancing Speed and Accuracy: Hybrid methods, such as those
this section, we shall summaries the remarkable advances made in the combining regression and segmentation-based approaches (e.g.,
domain of traffic sign detection over the last decade and examine po­ Spatial Attention-based Detection (SAD)), take advantage of the
tential future directions and challenges. speed of regression-based methods (such as YOLO and SSD) and the
detailed processing capabilities of segmentation-based approaches.
Conclusion and future trends This balance allows TSD models to achieve high accuracy while
maintaining the processing speed required for real-time detection,
In this paper, we present a comprehensive analysis of diverse ap­ which is crucial for autonomous driving and intelligent transport
plications of deep learning in the field of traffic sign detection. We systems.
highlight the unique advantages and challenges of regression, segmen­
tation, and hybrid approaches. Our study shows that while each These algorithms leverage the fast response of regression methods
approach has distinct application scenarios, hybrid algo­ and the precise details of segmentation methods, paving the way for the
rithms—especially those combining various complementary strat­ future development of traffic sign detection techniques.
egies—demonstrate significant potential for addressing TSD in complex The application of hybrid deep learning algorithms may be crucial
and dynamic environments. for the advancement of autonomous driving and intelligent transport
Regression-based detection algorithms, such as YOLO and SSD, boast systems, particularly in dealing with intricate and dynamic scenarios.
top-notch real-time performance and processing speed. Nevertheless, Furthermore, it is expected that these hybrid approaches will have a
more considerable impact on future TSD applications with the progress
in computational resources and algorithm optimization. Besides, The
challenges in TSD can be summarized as follows:

- Traditional TSD algorithms have limitations in detecting traffic signs

under challenging conditions or complex backgrounds. While these
algorithms have been effective under controlled environments and
specific cases, they struggle to generalize well when varying the
scales and orientations of traffic signs.
- Deep learning techniques have significantly improved traffic sign
detection capabilities. Regression-based approaches have proven
effective in detecting traffic signs with varying orientations, while
Fig. 18. Example of ROC and AUC [147]. segmentation-based algorithms, such as semantic segmentation,

14
H. Chen et al. Results in Engineering 24 (2024) 103553

instance segmentation, and hybrid approaches, have shown a based systems, resulting in improved performance and reliability of
promise in detecting the traffic signs under challenging conditions TSD systems. Simultaneously, it will be imperative to prioritize the
and complex backgrounds. advancement of TSD algorithms that can quickly and accurately
- Hybrid detection algorithms, which combine the strengths of process and analyze traffic signs in real-time while maintaining low
regression and segmentation-based approaches, can further improve latency.
the capabilities of TSD systems. However, developing algorithms - System Robustness and Energy Efficiency: As autonomous driving
that can accurately recognize and distinguish between a large technologies continue to develop, it is essential for TSD systems to be
number of traffic signs remains a complex task, raising questions more robust and capable of adapting to various lighting levels,
about reliability and efficiency in practical applications. Although adverse weather conditions, and occlusion. The achievement of long-
semantic and instance segmentation approaches have shown great term sustainable autonomous driving solutions requires the devel­
potential, they are still subjected to computational efficiency issues opment of new algorithms, implementation of data enhancement
and the challenge of effectively dealing with occlusions. techniques and utilization of advanced sensing technologies. Addi­
- The use of data augmentation and migration learning techniques can tionally, adopting more energy-efficient algorithms are needed.
enrich the training data and learn features from large datasets using - Knowledge transfer and data efficiency: Managing dissimilarities in
pre-trained models, while speeding up the training process, thus traffic signs across various regions and countries through transfer
improving the performance of TSD models. However, there is still the learning and domain adaptation is a significant area of research. It
problem of domain adaptation, and current techniques still struggle will involve finding solutions to synchronize the traffic signs across
to fully address the problem of unbalanced data and small samples. different regions globally. Meanwhile, improving the efficiency of
- Lighting Conditions and Occlusions: Crafting algorithms that can limited data through active learning strategies to reduce the time
accurately detect traffic signs in fluctuating lighting conditions and required for annotating large-scale datasets, will be a crucial
occlusions remains a significant challenge in the field. research focus.
- Algorithmic Complexity and Implementation Feasibility: it is still a - End-to-end embedded deployment and optimization: It is also
challenge to create an algorithm that balances speed and accuracy, necessary for future TSD research to enhance its deployment and
while remaining simple to implement in all situations. The above- optimization on embedded devices, especially on in-vehicle systems
listed challenges of TSD algorithms are all important to consider in cars or mobile devices. This necessitates the implementation of
when facing the future of Autonomous Driving roads towards algorithmic pruning and acceleration optimization for embedded
smarter, safer and more efficient. hardware (e.g. GPUs and AI accelerators) in order to ensure that the
model can maintain a high level of accuracy while meeting the
Our study indicates that TSD systems in the future should incorpo­ requisite real-time low-latency requirements. Furthermore, quanti­
rate algorithmic speed, accuracy, and flexibility to effectively adapt to zation and pruning techniques can be employed to reduce the size of
dynamic road environments. While this review focuses primarily on the model and the associated computational complexity, thus
algorithmic advancements, it is important to briefly acknowledge the enabling adaptation to scenarios in which the available hardware
role of hardware capabilities in the real-world implementation of TSD resources are limited.
systems. Key hardware considerations include camera quality, such as - Cross-Domain Adaptation and Multilingual Sign Recognition: Future
resolution, frame rate, and dynamic range, which significantly impact research could also address the adaptation of TSD systems to
the ability to capture clear and useful images of traffic signs, particularly accommodate various sign languages, scripts, and visual layouts
in adverse weather conditions or complex environments. Notably, used in different countries. The integration of cross-domain adap­
recent approaches involving the use of monocular and in-car cameras tation and multilingual text recognition in TSD systems would
have shown promising results in improving the accuracy and reliability enhance their global applicability, especially for international travel
of TSD. These techniques not only offer adaptability across varied en­ or logistics applications, where the vehicle may cross several regions
vironments but also demonstrate the integration of AR-based method­ with distinct traffic signs.
ologies for enhanced vehicular cooperation on the road. This indicates - Federated Learning for Distributed Model Training: The application
that the auch technologies could greatly contribute to real-world of federated learning in TSD systems can be explored to enable the
deployment, provided that computational constraints and camera use of distributed data for model training while maintaining data
quality limitations are adequately addressed. Additionally, the compu­ privacy. This approach could significantly improve the robustness
tational resources available on autonomous vehicles, including GPUs and generalizability of TSD models by using data from multiple
and AI accelerators, are crucial for running deep learning models effi­ sources without centralized data storage, which can be particularly
ciently in real-time, while maintaining low latency and optimizing en­ beneficial in training models across different geographic regions
ergy consumption. without privacy concerns.
In the future, there will likely be substantial advancements in mul­ - Human-Machine Interaction and User-Focused Design: Another
tiple dimensions within the field of TSD. Integrating multimodal data, emerging direction is to enhance the collaboration between the TSD
upgrading real-time processing techniques, optimizing algorithms for systems and human users, such as drivers or passengers. Human-
human-centered design and energy efficiency, will trigger improved centered design principles can improve the usability of these sys­
TSD research. TSD detection processing will evolve into an integrated tems and create user-friendly interfaces that provide accurate and
process, where the recognition phase not only offers more accurate re­ actionable information without overwhelming the driver. Research
sults but also helps eliminate non-textual distractions. Future research in this area can focus on enhancing the interpretability of TSD pre­
will prioritize combining TSD with real-time surveillance, multilingual dictions and designing mechanisms for seamless human-machine
text recognition, and mobile device detection for practical imple­ interaction to support driver awareness in real time.
mentations in real-world scenarios. There are several promising di­
rections for future development that can be considered, as follows: As research and development in the TSD field continue to progress,
we can expect even more advanced and efficient solutions to emerge,
- Technological advances and integration: TSD systems will further contributing to the growth of safer, smarter, and more sustainable
enhance the integration of multimodal data and real-time processing transportation systems worldwide.
capabilities, enabling the efficient detection and recognition of
traffic signs. This integration will include various data sources (such
as LiDAR, radar and other sensors) that work in tandem with vision-

15
H. Chen et al. Results in Engineering 24 (2024) 103553

CRediT authorship contribution statement [12] H.-Y. Lin, C.-C. Chang, V.L. Tran, J.-H. Shi, Improved traffic sign recognition for
in-car cameras, J. Chin. Inst. Eng. 43 (2020) 300–307, [Link]
02533839.2019.1708801.
Hui Chen: Conceptualization, Formal analysis, Methodology, Data [13] A.K. Dey, B. Goel, S. Chellappan, Context-driven detection of distracted driving
curation, Writing – original draft. Mohammed A.H. Ali: Conceptuali­ using images from in-car cameras, Internet Things 14 (2021) 100380.
zation, Formal analysis, Methodology, Data curation, Supervision, [14] X. Bai, P. Dong, Y. Huang, S. Kumari, H. Yu, Y. Ren, An AR-based meta vehicle
road cooperation testing systems: framework, components modeling, and an
Writing – review & editing, Funding acquisition, Project administration. implementation example, IEEE Internet Things J 11 (2024) 23460–23474,
Yusoff Nukman: Supervision, Writing – review & editing, Project [Link]
administration. Bushroa Abd Razak: Supervision, Writing – review & [15] J.-Y. Han, T.-H. Juan, T.-Y. Chuang, Traffic sign detection and positioning based
on monocular camera, J. Chin. Inst. Eng. 42 (2019) 757–769, [Link]
editing, Project administration. Sherzod Turaev: Writing – review & 10.1080/02533839.2019.1660220.
editing, Funding acquisition, Project administration. YiHan Chen: [16] C. Liu, S. Li, F. Chang, Y. Wang, Machine vision based traffic sign detection
Methodology, Data curation. Shikai Zhang: Methodology, Data cura­ methods: review, analyses and perspectives, IEEE Access 7 (2019) 86578–86596,
[Link]
tion. Zhiwei Huang: Methodology, Data curation. Zhenya Wang: [17] M. Swathi, K.V. Suresh, Automatic traffic sign detection and recognition: a
Methodology, Data curation. Rawad Abdulghafor: Writing – review & review, in: 2017 Int. Conf. Algorithms Methodol. Models Appl. Emerg. Technol.
editing, Funding acquisition, Project administration. ICAMMAET, IEEE, 2017, pp. 1–6, [Link]
ICAMMAET.2017.8186650.
[18] Z. Liu, D. Li, S.S. Ge, F. Tian, Small traffic sign detection from large image, Appl.
Declaration of competing interest Intell. 50 (2020) 1–13, [Link]
[19] Y. Tian, J. Gelernter, X. Wang, J. Li, Y. Yu, Traffic sign detection using a multi-
scale recurrent attention network, IEEE Trans. Intell. Transp. Syst. 20 (2019)
The authors declare that they have no known competing financial 4466–4475, [Link]
interests or personal relationships that could have appeared to influence [20] X. Chen, K. Yang, Y. Xu, K. Li, Top-100 highest-cited original articles in
the work reported in this paper. inflammatory bowel disease: a bibliometric analysis, Medicine (Baltimore) 98
(2019) e15718, [Link]
[21] M.F. Perazzo, A.L.C. Otoni, M.S. Costa, A.F. Granville-Granville, S.M. Paiva, P.
Acknowledgements A. Martins-Júnior, The top 100 most-cited papers in Paediatric dentistry journals:
a bibliometric analysis, Int. J. Paediatr. Dent. 29 (2019) 692–711, [Link]
org/10.1111/ipd.12563.
Authors would like to thank University of Malaya and Ministry of
[22] S.B. Wali, M.A. Abdullah, M.A. Hannan, A. Hussain, S.A. Samad, P.J. Ker, M.
High Education-Malaysia for supporting this work via Fundemental B. Mansor, Vision-based traffic sign detection and recognition systems: current
Research Grant Scheme (FRGS/1/2023/TK10/UM/02/3 and GPF020A- trends and challenges, Sensors 19 (2019) 2093, [Link]
2023). The authors also thank Sandooq Al Watan, United Arab Emirates, s19092093.
[23] S. Houben, J. Stallkamp, J. Salmen, M. Schlipsing, C. Igel, Detection of traffic
for supporting this work through the SWARD Research Grant signs in real-world images: the german traffic sign detection benchmark, in: 2013
G00003606 and United Arab Emirates University through the UAEU Int. Jt. Conf. Neural Netw. IJCNN, Ieee, 2013: pp. 1–8. [Link]
Strategic Research Grant G00003676 via the Big Data Analytics Center. /IJCNN.2013.6706807.
[24] Z. Zhu, D. Liang, S. Zhang, X. Huang, B. Li, S. Hu, Traffic-sign detection and
classification in the wild, in: Proc. IEEE Conf. Comput. Vis. Pattern Recognit.,
Data availability 2016: pp. 2110–2118.
[25] Drive in Malaysia – Malaysia’s first ever traffic rules and test website, (n.d.).
[Link] May 6, 2023).
No data was used for the research described in the article. [26] M. Mathias, R. Timofte, R. Benenson, L. Van Gool, Traffic sign recognition—how
far are we from the solution?, in: 2013 Int. Jt. Conf. Neural Netw. IJCNN IEEE,
References 2013, pp. 1–8.
[27] F. Larsson, M. Felsberg, Using Fourier descriptors and spatial models for traffic
sign recognition, in: Image Anal. 17th Scand. Conf. SCIA 2011 Ystad Swed. May
[1] R. Ratajczak, C.F. Crispim-Junior, E. Faure, B. Fervers, L. Tougne, Automatic land
2011 Proc. 17, Springer, 2011, pp. 238–249, [Link]
cover reconstruction from historical aerial images: An evaluation of features
21227-7_23.
extraction and classification algorithms, IEEE Trans. Image Process. 28 (2019)
[28] V. Dalborgo, T.B. Murari, V.S. Madureira, J.G.L. Moraes, V.M.O. Bezerra, F.
3357–3371, [Link]
Q. Santos, A. Silva, R.L. Monteiro, Traffic sign recognition with deep learning:
[2] D. Tabernik, D. Skočaj, Deep learning for large-scale traffic-sign detection and
vegetation occlusion detection in brazilian environments, Sensors 23 (2023)
recognition, IEEE Trans. Intell. Transp. Syst. 21 (2019) 1427–1440, [Link]
5919.
org/10.1109/TITS.2019.2913588.
[29] A. Alam, Z.A. Jaffery, Indian traffic sign detection and recognition, Int. J. Intell.
[3] T. Yang, X. Long, A.K. Sangaiah, Z. Zheng, C. Tong, Deep detection network for
Transp. Syst. Res. 18 (2020) 98–112.
real-life traffic sign in vehicular networks, Comput. Netw. 136 (2018) 95–104,
[30] H. Gómez-Moreno, S. Maldonado-Bascón, P. Gil-Jiménez, S. Lafuente-Arroyo,
[Link]
Goal evaluation of segmentation algorithms for traffic sign recognition, IEEE
[4] Y. Berhanu, D. Schröder, B.T. Wodajo, E. Alemayehu, Machine learning for
Trans. Intell. Transp. Syst. 11 (2010) 917–930.
predictions of road traffic accidents and spatial network analysis for safe routing
[31] T.T.A. Cu, X.C. Vuong, C.T. Nguyen, T.T. Vu, T.A. Nguyen, Detection of
on accident and congestion-prone road networks, Results Eng 23 (2024) 102737,
vietnamese traffic danger and warning signs via deep learning, J. Eng. Sci.
[Link]
Technol. 19 (2024) 133–145.
[5] G. Cui, L. Zhang, Improved faster region convolutional neural network algorithm
[32] Y. Yuan, Z. Xiong, Q. Wang, VSSA-NET: Vertical spatial sequence attention
for UAV target detection in complex environment, Results Eng 23 (2024) 102487,
network for traffic sign detection, IEEE Trans. Image Process. 28 (2019)
[Link]
3423–3434, [Link]
[6] M. Kalinsky, M. Sykora, J. Markova, P. Maňas, Nonlinear dynamic finite element
[33] D. Temel, G. Kwon, M. Prabhushankar, G. AlRegib, CURE-TSR: challenging unreal
analysis of vehicle impacts into road restraint systems, Results Eng 23 (2024)
and real environments for traffic sign recognition, ArXiv Prepr (2017).
102726, [Link]
ArXiv171202463.
[7] M.S. Mohammed, A.M. Abduljabar, M.M. Faisal, B.M. Mahmmod, S.
[34] K. Bayoudh, F. Hamdaoui, A. Mtibaa, Transfer learning based hybrid 2D-3D CNN
H. Abdulhussain, W. Khan, P. Liatsis, A. Hussain, Low-cost autonomous car level
for traffic sign recognition and semantic road detection applied in advanced
2: design and implementation for conventional vehicles, Results Eng 17 (2023)
driver assistance systems, Appl. Intell. 51 (2021) 124–142, [Link]
100969, [Link]
10.1007/s10489-020-01801-5.
[8] J. Nan, Z. Ge, X. Ye, A.F. Burke, J. Zhao, Model predictive control for autonomous
[35] D. Horn, S. Houben, Fully automated traffic sign substitution in real-world images
vehicle path tracking through optimized kinematics, Results Eng. (2024) 103123,
for large-scale data augmentation, in: 2020 IEEE Intell. Veh. Symp. IV, IEEE,
[Link]
2020, pp. 465–471, [Link]
[9] H.A. Neamah, E. Donát, P. Korondi, Optimizing autonomous navigation in
[36] M.Z. Abedin, P. Dhar, K. Deb, Traffic sign recognition using hybrid features
unknown environments: a novel trap avoiding vector field histogram algorithm
descriptor and artificial neural network classifier, in: 2016 19th Int. Conf.
VFH+T, Results Eng. 23 (2024) 102625, [Link]
Comput. Inf. Technol. ICCIT, IEEE, 2016, pp. 457–462, [Link]
rineng.2024.102625.
ICCITECHN.2016.7860241.
[10] N. Promkaew, S. Thammawiset, P. Srisan, P. Sanitchon, T. Tummawai, S.
[37] B. Cyganek, Color image segmentation with support vector machines:
Sukpancharoen, Development of metaheuristic algorithms for efficient path
applications to road signs detection, Int. J. Neural Syst. 18 (2008) 339–345,
planning of autonomous mobile robots in indoor environments, Results Eng. 22
[Link]
(2024) 102280. [Link]
[38] dge detection Kamal, T.I. Tonmoy, S. Das, M.K. Hasan, Automatic traffic sign
[11] K. Vinoth, P. Sasikumar, Lightweight object detection in low light: Pixel-wise
detection and recognition using SegU-Net and a modified Tversky loss function
depth refinement and TensorRT optimization, Results Eng. 23 (2024) 102510,
[Link]

16
H. Chen et al. Results in Engineering 24 (2024) 103553

with L1-constraint, IEEE Trans. Intell. Transp. Syst. 21 (2019) 1467–1479, [65] A. Muhammad, M.A.H. Ali, S. Turaev, I.H. Shanono, F. Hujainah, M.N.M. Zubir,
[Link] M.K. Faiz, E.R.M. Faizal, R. Abdulghafor, Novel algorithm for mobile robot path
[39] R. Girshick, J. Donahue, T. Darrell, J. Malik, Rich feature hierarchies for accurate planning in constrained environment, Comput. Mater. Contin. 71 (2022)
object detection and semantic segmentation, in: Proc. IEEE Conf. Comput. Vis. 2697–2719, [Link]
Pattern Recognit, 2014, pp. 580–587. [66] H. Wang, S. Lou, J. Jing, Y. Wang, W. Liu, T. Liu, The EBS-A* algorithm: an
[40] M. Hashimoto, F. Oba, T. Tomiie, Mobile robot localization using color signboard, improved A* algorithm for path planning, PLoS ONE 17 (2022) 1–27, [Link]
Mechatronics 9 (1999) 633–656, [Link] org/10.1371/[Link].0263841.
00012-4. [67] W. Min, R. Liu, D. He, Q. Han, Q. Wei, Q. Wang, Traffic sign recognition based on
[41] S. Saha, N. Chakraborty, S. Kundu, S. Paul, A.F. Mollah, S. Basu, R. Sarkar, Multi- semantic scene understanding and structural traffic sign location, IEEE Trans.
lingual scene text detection and language identification, Pattern Recognit. Lett. Intell. Transp. Syst. 23 (2022) 15794–15807, [Link]
138 (2020) 16–22, [Link] TITS.2022.3145467.
[42] S. Salti, A. Petrelli, F. Tombari, N. Fioraio, L. Di Stefano, Traffic sign detection via [68] Y. Liu, J. Peng, J.-H. Xue, Y. Chen, Z.-H. Fu, TSingNet: Scale-aware and context-
interest region extraction, Pattern Recognit 48 (2015) 1039–1049, [Link] rich feature learning for traffic sign detection and recognition in the wild,
org/10.1016/[Link].2014.05.017. Neurocomputing 447 (2021) 10–22, [Link]
[43] P.K. Perepu, Deep learning for detection of text polarity in natural scene images, neucom.2021.03.049.
Neurocomputing 431 (2021) 1–6, [Link] [69] H.S. Lee, K. Kim, Simultaneous traffic sign detection and boundary estimation
neucom.2020.12.054. using convolutional neural network, IEEE Trans. Intell. Transp. Syst. 19 (2018)
[44] Á. Arcos-García, J.A. Álvarez-García, L.M. Soria-Morillo, Evaluation of deep 1652–1663, [Link]
neural networks for traffic sign detection systems, Neurocomputing 316 (2018) [70] L. He, F. Lan, C. Zhou, Y. Ye, W. Zhang, B. Chen, J. Pan, A feature-enhanced
332–344, [Link] hybrid attention network for traffic sign recognition in real scenes, IET IMAGE
[45] Z. Liu, T. Huang, B. Li, X. Chen, X. Wang, X. Bai, EPNet++: cascade bi-directional Process 18 (2024) 2064–2077, [Link]
fusion for multi-modal 3D object detection, IEEE Trans. Pattern Anal. Mach. [71] P. Liu, Z. Xie, T. Li, UCN-YOLOv5: traffic sign object detection algorithm based on
Intell. (2022), [Link] deep learning, IEEE Access 11 (2023) 110039–110050, [Link]
[46] Z. Nadeem, Z. Khan, U. Mir, U.I. Mir, S. Khan, H. Nadeem, J. Sultan, Pakistani ACCESS.2023.3322371.
traffic-sign recognition using transfer learning, Multimed. Tools Appl. 81 (2022) [72] A.J. Prakash, S. Sruthy, Enhancing traffic sign recognition (TSR) by classifying
8429–8449, [Link] deep learning models to promote road safety, Signal Image Video Process 18
[47] J. Cao, C. Song, S. Peng, F. Xiao, S. Song, Improved traffic sign detection and (2024) 4713–4729, [Link]
recognition algorithm for intelligent vehicles, Sensors 19 (2019) 4021, https:// [73] B. Chen, X. Fan, MSGC-YOLO: an improved lightweight traffic sign detection
[Link]/10.3390/s19184021. model under snow conditions, Mathematics 12 (2024) 1539, [Link]
[48] J. Chung, S. Park, D. Pae, H. Choi, M. Lim, Feature-selection-based attentional- 10.3390/math12101539.
Deconvolution detector for German traffic sign detection benchmark, Electronics [74] Y. Jin, Y. Fu, W. Wang, J. Guo, C. Ren, X. Xiang, Multi-feature fusion and
12 (2023) 725, [Link] enhancement single shot detector for traffic sign recognition, IEEE Access 8
[49] A. Latif, A. Rasheed, U. Sajid, J. Ahmed, N. Ali, N.I. Ratyal, B. Zafar, S.H. Dar, (2020) 38931–38940, [Link]
M. Sajid, T. Khalil, Content-based image retrieval and feature extraction: a [75] S. Saxena, S. Dey, M. Shah, S. Gupta, Traffic sign detection in unconstrained
comprehensive review, Math. Probl. Eng. (2019) 2019, [Link] environment using improved YOLOv4, Expert Syst. Appl. 238 (2024) 121836,
2019/9658350. [Link]
[50] M.A.A. Sheikh, A. Kole, T. Maity, Traffic sign detection and classification using [76] C. Sun, M. Wen, K. Zhang, P. Meng, R. Cui, Traffic sign detection algorithm based
colour feature and neural network, in: 2016 Int. Conf. Intell. Control Power on feature expression enhancement, Multimed. Tools Appl. 80 (2021)
Instrum. ICICPI, IEEE, 2016, pp. 307–311, [Link] 33593–33614, [Link]
ICICPI.2016.7859723. [77] J. Wang, Y. Chen, Z. Dong, M. Gao, Improved YOLOv5 network for real-time
[51] J. Zhang, X. Liu, W. Liao, X. Li, Deep-learning generation of POI data with scene multi-scale traffic sign detection, Neural Comput. Appl. 35 (2023) 7853–7865,
images, ISPRS J. Photogramm. Remote Sens. 188 (2022) 201–219, [Link] [Link]
org/10.1016/[Link].2022.04.004. [78] X. Wang, Y. Tian, K. Zheng, C. Liu, C2Net-YOLOv5: a bidirectional Res2Net-based
[52] R. Belaroussi, P. Foucher, J.-P. Tarel, B. Soheilian, P. Charbonnier, N. Paparoditis, traffic sign detection algorithm, Comput. Mater. Contin. 77 (2023) 1949–1965,
Road sign detection in images: a case study, in: 2010 20th Int. Conf. Pattern [Link]
Recognit, IEEE, 2010, pp. 484–488, [Link] [79] J. Wu, S. Liao, Traffic sign detection based on SSD combined with receptive field
[53] C. Grigorescu, N. Petkov, M.A. Westenberg, Contour detection based on module and path aggregation network, Comput. Intell. Neurosci. (2022) 2022.
nonclassical receptive field inhibition, IEEE Trans. Image Process. 12 (2003) [80] X. Zhang, Y. Tian, Traffic sign detection algorithm based on improved YOLOv8s,
729–739, [Link] Eng. Lett. 32 (2024) 168–178.
[54] M. Shi, H. Wu, H. Fleyeh, Support vector machines for traffic signs recognition, [81] C. Dewi, R.-C. Chen, H. Yu, Weight analysis for various prohibitory sign detection
in: 2008 IEEE Int. Jt. Conf. Neural NetwIEEE World Congr. Comput. Intell., IEEE, and recognition using deep learning, Multimed. Tools Appl. 79 (2020)
2008, pp. 3820–3827. 32897–32915, [Link]
[55] Y. Jiang, S. Zhou, Y. Jiang, J. Gong, G. Xiong, H. Chen, Traffic sign recognition [82] Y. Zhu, W.Q. Yan, Traffic sign recognition based on deep learning, Multimed.
using ridge regression and Otsu method, in: 2011 IEEE Intell. Veh. Symp. IV, Tools Appl. 81 (2022) 17779–17791, [Link]
IEEE, 2011, pp. 613–618, [Link] 12163-0.
[56] C.-Y. Fang, S.-W. Chen, C.-S. Fuh, Road-sign detection and tracking, IEEE Trans. [83] J. Xia, M. Li, W. Liu, X. Chen, DSRA-DETR: an improved DETR for Multiscale
Veh. Technol. 52 (2003) 1329–1341, [Link] traffic sign detection, Sustainability 15 (2023) 10862, [Link]
TVT.2003.810999. su151410862.
[57] X. Yuan, X. Hao, H. Chen, X. Wei, Traffic sign recognition based on a context- [84] C. Han, G. Gao, Y. Zhang, Real-time small traffic sign detection with revised
aware scale-invariant feature transform approach, J. Electron. Imag. 22 (2013) faster-RCNN, Multimed. Tools Appl. 78 (2019) 13263–13278, [Link]
041105, [Link] –041105. 10.1007/s11042-018-6428-0.
[58] T.T. Dang, H.Y. Ngan, W. Liu, Distance-based k-nearest neighbors outlier [85] R.C. Rodríguez, C.M. Carlos, O.O.V. Villegas, V.G.C. Sánchez, H. de J.O.
detection method in large-scale traffic data, in: 2015 IEEE Int. Conf. Digit. Signal Domínguez, Mexican traffic sign detection and classification using deep learning,
Process. DSP, IEEE, 2015, pp. 507–510, [Link] Expert Syst. Appl. 202 (2022) 117247.
ICDSP.2015.7251924. [86] S.K. Satti, P. Maddula, N.V. Ravipati, Unified approach for detecting traffic signs
[59] A. Hechri, A. Mtibaa, Two-stage traffic sign detection and recognition based on and potholes on Indian roads, J. King Saud Univ.-Comput. Inf. Sci. (2021).
SVM and convolutional neural networks, IET Image Process 14 (2020) 939–946, [87] L. Huang, H. Wang, J. Zeng, S. Zhang, L. Cao, J. Yan, H. Li, Geometric-aware
[Link] pretraining for vision-centric 3D object detection, (2023). [Link]
[60] C.G. Kiran, L.V. Prabhu, K. Rajeev, Traffic sign detection and pattern recognition 48550/arXiv.2304.03105.
using support vector machine, in: 2009 Seventh Int. Conf. Adv. Pattern Recognit, [88] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, A.C. Berg, Comput.
IEEE, 2009, pp. 87–90, [Link] Vis. – ECCV 2016, in: B. Leibe, J. Matas, N. Sebe, M. Welling (Eds.), SSD: Single
[61] S. Maldonado-Bascón, S. Lafuente-Arroyo, P. Gil-Jimenez, H. Gómez-Moreno, Shot MultiBox Detector, Springer International Publishing, Cham, 2016,
F. López-Ferreras, Road-sign detection and recognition based on support vector pp. 21–37, [Link]
machines, IEEE Trans. Intell. Transp. Syst. 8 (2007) 264–278, [Link] [89] J. Redmon, A. Farhadi, YOLOv3: An Incremental Improvement, (2018). htt
10.1109/TITS.2007.895311. ps://[Link]/10.48550/arXiv.1804.02767.
[62] Y. Zhu, M. Liao, M. Yang, W. Liu, Cascaded segmentation-detection networks for [90] R. Faster, Towards real-time object detection with region proposal networks, Adv.
text-based traffic sign detection, IEEE Trans. Intell. Transp. Syst. 19 (2017) Neural Inf. Process. Syst. 9199 (2015) 2969239–2969250.
209–219, [Link] [91] V.-D. Hoang, M.-H. Le, T.T. Tran, V.-H. Pham, Improving traffic signs recognition
[63] Y. Perez-Perez, M. Golparvar-Fard, K. El-Rayes, Segmentation of point clouds via based region proposal and deep neural networks, in: N.T. Nguyen, D.H. Hoang,
joint semantic and geometric features for 3D modeling of the built environment, T.-P. Hong, H. Pham, B. Trawiński (Eds.), Improving traffic signs recognition
Autom. Constr. 125 (2021) 103584, [Link] based region proposal and deep neural networks, Intell. Inf. Database Syst. (2018)
autcon.2021.103584. 604–613, [Link]
[64] R. Yazdan, M. Varshosaz, Improving traffic sign recognition results in urban areas [92] S. Du, W. Pan, N. Li, S. Dai, B. Xu, H. Liu, C. Xu, X. Li, TSD-YOLO: small traffic
by overcoming the impact of scale and rotation, ISPRS J. Photogramm. Remote sign detection based on improved YOLO v8, IET Image Process 18 (2024)
Sens. 171 (2021) 18–35, [Link] 2884–2898, [Link]

17
H. Chen et al. Results in Engineering 24 (2024) 103553

[93] C. Lyu, X. Fan, Z. Qiu, J. Chen, J. Lin, C. Dong, Efficientdet based visial perception [119] H.-Y. Lin, S.-Y. Lin, K.-C. Tu, Traffic light detection and recognition using a two-
for autonomous driving, in: 2023 8th Int. Conf. Cloud Comput. Big Data Anal. stage framework from individual signal bulb identification, IEEE Access 12
ICCCBDA, IEEE, Chengdu, China, 2023, pp. 443–447, [Link] (2024) 132279–132289, [Link]
ICCCBDA56900.2023.10154715. [120] Z. Li, H. Chen, B. Biggio, Y. He, H. Cai, F. Roli, L. Xie, Toward effective traffic sign
[94] X. Qian, Y. Liu, Y. Yang, MGPAN: mask guided pixel aggregation network, in: detection via two-stage fusion neural networks, IEEE Trans. Intell. Transp. Syst.
2020 IEEE Int. Conf. Image Process, ICIP, 2020, pp. 1981–1985, [Link] 25 (2024) 8283–8294, [Link]
10.1109/ICIP40778.2020.9190897. [121] C. Shorten, T.M. Khoshgoftaar, A survey on image data augmentation for deep
[95] Y. Zhu, J. Chen, L. Liang, Z. Kuang, L. Jin, W. Zhang, Fourier contour embedding learning, J. Big Data 6 (2019) 1–48, [Link]
for arbitrary-shaped text detection, in: 2021: pp. 3123–3131. [Link] 0.
[Link]/content/CVPR2021/html/Zhu_Fourier_Contour_Embedding_for_Arbit [122] L. Jöckel, M. Kläs, S. Martínez-Fernández, Safe traffic sign recognition through
rary-Shaped_Text_Detection_CVPR_2021_paper.html (accessed May 5, 2023). data augmentation for autonomous vehicles software, in: 2019 IEEE 19th Int.
[96] Y. Zhu, C. Zhang, D. Zhou, X. Wang, X. Bai, W. Liu, Traffic sign detection and Conf. Softw. Qual. Reliab. Secur. Companion QRS-C, IEEE, 2019, pp. 540–541,
recognition using fully convolutional network guided proposals, Neurocomputing [Link]
214 (2016) 758–766. [123] F. Mumcu, Y. Yilmaz, Fast and lightweight vision-language model for adversarial
[97] X. Li, Z. Xie, X. Deng, Y. Wu, Y. Pi, Traffic sign detection based on improved faster traffic sign detection, ELECTRONICS 13 (2024) 2172, [Link]
R-CNN for autonomous driving, J. Supercomput. 78 (2022) 7982–8002, https:// electronics13112172.
[Link]/10.1007/s11227-021-04230-4. [124] N. Soufi, M. Valdenegro-Toro, Data augmentation with symbolic-to-real image
[98] O. Ronneberger, P. Fischer, T. Brox, U-Net: convolutional networks for translation GANs for traffic sign recognition, ArXiv Prepr (2019)
biomedical image segmentation, in: N. Navab, J. Hornegger, W.M. Wells, A. ArXiv190712902.
F. Frangi (Eds.), Med. Image Comput. Comput.-Assist. Interv. – MICCAI 2015, [125] Y. Wu, Z. Li, Y. Chen, K. Nai, J. Yuan, Real-time traffic sign detection and
Springer International Publishing, Cham, 2015, pp. 234–241, [Link] classification towards real traffic scene, Multimed. Tools Appl. 79 (2020)
10.1007/978-3-319-24574-4_28. 18201–18219, [Link]
[99] S. Zhao, Z. Gong, D. Zhao, Traffic signs and markings recognition based on [126] S. Zhao, Z. Gong, D. Zhao, Traffic signs andmarkings recognition based on
lightweight convolutional neural network, Vis. Comput. (2023) 1–12. lightweight convolutional neural network, Vis. Comput. 40 (2024) 559–570,
[100] L.-C. Chen, Y. Zhu, G. Papandreou, F. Schroff, H. Adam, Encoder-decoder with [Link]
atrous separable convolution for semantic image segmentation, in: Proc. Eur. [127] Q. Li, J. Zheng, W. Tan, X. Wang, Y. Zhao, Traffic sign detection: appropriate data
Conf. Comput. Vis. ECCV, 2018, pp. 801–818. augmentation method from the perspective of frequency domain, Math. Probl.
[101] Y. Saadna, A. Behloul, An overview of traffic sign detection and classification Eng. 2022 (2022) 1–11, [Link]
methods, Int. J. Multimed. Inf. Retr. 6 (2017) 193–210, [Link] [128] J. Zhang, Z. Xie, J. Sun, X. Zou, J. Wang, A cascaded R-CNN with multiscale
s13735-017-0129-8. attention and imbalanced samples for traffic sign detection, IEEE Access 8 (2020)
[102] M. Soilán, B. Riveiro, J. Martinez-Sanchez, P. Arias, Traffic sign detection in MLS 29742–29754, [Link]
acquired point clouds for geometric and image-based semantic inventory, ISPRS [129] A. Mannan, K. Javed, A.U. Rehman, H.A. Babri, S.K. Noon, Classification of
J. Photogramm. Remote Sens. 114 (2016) 92–101, [Link] degraded traffic signs using flexible mixture model and transfer learning, IEEE
isprsjprs.2016.01.019. Access 7 (2019) 148800–148813, [Link]
[103] S. Liu, L. Qi, H. Qin, J. Shi, J. Jia, Path aggregation network for instance ACCESS.2019.2947069.
segmentation, in: 2018 IEEECVF Conf. Comput. Vis. Pattern Recognit., IEEE, Salt [130] P. Kora, C.P. Ooi, O. Faust, U. Raghavendra, A. Gudigar, W.Y. Chan,
Lake City, UT, 2018, pp. 8759–8768, [Link] K. Meenakshi, K. Swaraja, P. Plawiak, U.R. Acharya, Transfer learning techniques
CVPR.2018.00913. for medical image analysis: a review, Biocybern. Biomed. Eng. 42 (2022) 79–107,
[104] D. Deng, H. Liu, X. Li, D. Cai, Pixellink: detecting scene text via instance [Link]
segmentation, in: Proc. AAAI Conf. Artif. Intell., 2018, [Link] [131] Y. Liu, Q. Qian, H. Zhang, J. Li, Y. Zhong, N.N. Xiong, Application of sustainable
aaai.v32i1.12269. blockchain technology in the internet of vehicles: innovation in traffic sign
[105] J. Liu, X. Liu, J. Sheng, D. Liang, X. Li, Q. Liu, Pyramid mask text detector, (2019). detection systems, Sustainability 16 (2024) 171, [Link]
[Link] (accessed April 4, 2023). su16010171.
[106] C. Ai, Y.J. Tsai, Hybrid active contour–incorporated sign detection algorithm, [132] J. Lu, V. Behbood, P. Hao, H. Zuo, S. Xue, G. Zhang, Transfer learning using
J. Comput. Civ. Eng. 26 (2012) 28–36, [Link] computational intelligence: a survey, Knowl.-Based Syst 80 (2015) 14–23,
5487.0000110. [Link]
[107] V. Balali, A. Jahangiri, S.G. Machiani, Multi-class US traffic signs 3D recognition [133] K. Weiss, T.M. Khoshgoftaar, D. Wang, A survey of transfer learning, J. Big Data 3
and localization via image-based point cloud model using color candidate (2016) 1–40, [Link]
extraction and texture-based recognition, Adv. Eng. Inform. 32 (2017) 263–274, [134] H. Zhao, X. Peng, S. Wang, J.-B. Li, J.-S. Pan, X. Su, X. Liu, Improved object
[Link] detection method for unmanned driving based on Transformers, Front.
[108] B. Riveiro, L. Díaz-Vilariño, B. Conde-Carnero, M. Soilán, P. Arias, Automatic Neurorobotics 18 (2024) 1342126, [Link]
segmentation and shape-based classification of retro-reflective traffic signs from fnbot.2024.1342126.
mobile LiDAR data, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 9 (2015) [135] H. Luo, Y. Yang, B. Tong, F. Wu, B. Fan, Traffic sign recognition using a multi-task
295–303, [Link] convolutional neural network, IEEE Trans. Intell. Transp. Syst. 19 (2018)
[109] Fusion of linear and non-linear dimensionality reduction techniques for feature 1100–1111, [Link]
reduction in LSTM-based intrusion detection system, Appl. Soft Comput. 154 [136] R.-C. Chen, C. Dewi, S.-W. Huang, R.E. Caraka, Selecting critical features for data
(2024) 111378, [Link] classification based on machine learning methods, J. Big Data 7 (2020) 52,
[110] P. Polewski, J. Shelton, W. Yao, M. Heurich, Instance segmentation of fallen trees [Link]
in aerial color infrared imagery using active multi-contour evolution with fully [137] C. Dewi, R.-C. Chen, Human activity recognition based on evolution of features
convolutional network-based intensity priors, ISPRS J. Photogramm. Remote selection and random forest, in: 2019 IEEE Int. Conf. Syst. Man Cybern, SMC,
Sens. 178 (2021) 297–313, [Link] 2019, pp. 2496–2501, [Link]
[111] C. Dewi, R.-C. Chen, X. Jiang, H. Yu, Deep convolutional neural network for [138] H. Kang, C. Chen, Fast implementation of real-time fruit detection in apple
enhancing traffic sign recognition developed on Yolo V4, Multimed. Tools Appl. orchards using deep learning, Comput. Electron. Agric. 168 (2020) 105108,
81 (2022) 37821–37845, [Link] [Link]
[112] S. Liu, T. Cai, X. Tang, Y. Zhang, C. Wang, Visual recognition of traffic signs in [139] R. Shi, T. Li, Y. Yamaguchi, An attribution-based pruning method for real-time
natural scenes based on improved RetinaNet, Entropy 24 (2022) 112, [Link] mango detection with YOLO network, Comput. Electron. Agric. 169 (2020)
org/10.3390/e24010112. 105214, [Link]
[113] H. Visaria, S. Chaube, F. John, A. Khubalkar, TSRSY-traffic sign recognition [140] Y. Tian, G. Yang, Z. Wang, H. Wang, E. Li, Z. Liang, Apple detection during
system using deep learning, in: 2022 2nd Int. Conf. Intell. Technol. CONIT, IEEE, different growth stages in orchards using the improved YOLO-V3 model, Comput.
2022, pp. 1–6, [Link] Electron. Agric. 157 (2019) 417–426, [Link]
[114] X. Zhang, H. Li, F. Meng, Z. Song, L. Xu, Segmenting beyond the bounding box for compag.2019.01.012.
instance segmentation, IEEE Trans. Circuits Syst. Video Technol. 32 (2022) [141] K. Geng, G. Yin, Using deep learning in infrared images to enable human gesture
704–714, [Link] recognition for autonomous vehicles, IEEE Access 8 (2020) 88227–88240,
[115] X. Qian, S. Lin, G. Cheng, X. Yao, H. Ren, W. Wang, Object detection in remote [Link]
sensing images based on improved bounding box regression and multi-level [142] J. Liu, Y. Huang, J. Peng, J. Yao, L. Wang, Fast object detection at constrained
features fusion, Remote Sens 12 (2020) 143, [Link] energy, IEEE Trans. Emerg. Top. Comput. 6 (2018) 409–416, [Link]
rs12010143. 10.1109/TETC.2016.2577538.
[116] G. Zeng, Z. Wu, L. Xu, Y. Liang, Efficient vision transformer YOLOv5 for accurate [143] C. Dewi, R.-C. Chen, Y.-T. Liu, H. Yu, Various generative adversarial networks
and fast traffic sign detection, Electronics 13 (2024) 880, [Link] model for synthetic prohibitory sign image generation, Appl. Sci. 11 (2021) 2913,
10.3390/electronics13050880. [Link]
[117] J. Hou, H. Zeng, L. Cai, J. Zhu, J. Cao, J. Hou, Handwritten numeral recognition [144] C. Dewi, R.-C. Chen, Random forest and support vector machine on features
using multi-task learning, in: 2017 Int. Symp. Intell. Signal Process. Commun. selection for regression analysis, (2019). [Link]
Syst. ISPACS, IEEE, 2017, pp. 155–158. 6.2027.
[118] F. Ren, H. Zhou, L. Yang, F. Liu, X. He, ADPNet: attention based dual path
network for lane detection, J. Vis. Commun. Image Represent. 87 (2022) 103574,
[Link]

18
H. Chen et al. Results in Engineering 24 (2024) 103553

[145] T. Liang, H. Bao, W. Pan, F. Pan, Traffic 33Sign detection via improved sparse R- [147] C. Chang, Multiparameter receiver operating characteristic analysis for signal
CNN for autonomous vehicles, J. Adv. Transp. 2022 (2022) 1–16, [Link] detection and classification, IEEE Sens. J. 10 (2010) 423–442, [Link]
10.1155/2022/3825532. 10.1109/JSEN.2009.2038120.
[146] J. Redmon, A. Farhadi, Yolov3: an incremental improvement, Comput. Vis,
Pattern Recognit 18 (2018) 1804–2767.

19