Evolution of Automated Pothole Detection and

Multi-Modal Surface Analysis*

Abstract—Road infrastructure has a direct impact on both handheld tools. This approach is inherently slow, varies con-
public safety and the overall economic conditions of a region. siderably from one inspector to another, and puts workers
Conventional road inspection methods, which are largely depend at risk when surveys are conducted on busy highways. The
on manual labor and subjective visual judgment, are becom-
ing increasingly insufficient given the growing scale of traffic subjective nature of such assessments also makes it hard to
and the need of real-time, automated monitoring. This paper do consistent trend analysis or properly prioritize repairs.
presents a review of automated pothole detection technologies As urban populations expand and traffic volumes increase,
that were developed between 2020 and 2026, and trace how the the inadequacies of these traditional methods have become
field moved from basic image classifiers toward more complex, increasingly hard to ignore.
multi-modal sensing pipelines. We look at three main detection
approaches, namely IMU-based vibration sensing, vision-driven
deep learning, and high-accuracy 3D surface reconstruction
using LiDAR. The paper also covers key developments in model
architecture like NMS-free YOLOv10 and Vision Transformers
(ViT), as well as newer strategies for privacy-preserving training
via Federated Learning. Based on the studies reviewed, vibration-
based methods appear to be more practical for wide-scale
deployment, but for precise volumetric measurements, LiDAR-
based 3D reconstruction is still considered the gold standard.
The review ends by pointing out some unresolved challenges
and future directions, particularly around edge computing and
decentralized learning for road monitoring systems.
Index Terms—pothole detection, deep learning, LiDAR,
YOLOv10, federated learning, inertial measurement unit, road
surface analysis, Vision Transformers, NPD-GAN, 3D reconstruc-
tion

I. I NTRODUCTION
Fig. 1.
Road surface quality is among the most obvious signs of
a city’s infrastructure condition, yet it is often one of the From 2020 onward, the field has undergone quite a signifi-
most neglected aspects of urban planning. Potholes, which cant transformation in terms of how road defects are detected,
are essentially localized failures in the pavement caused by measured, and reported. Progress in embedded sensor hard-
a combination of mechanical load, water penetration, and ware, mobile computing, deep learning, and computer vision
temperature variation, remain a persistent problem for trans- has enabled a new generation of automated systems that work
port authorities in both developed and developing regions. continuously and at large scale. Smartphones and dashcams,
The impact of such defects goes well beyond inconvenience which are now widely available and relatively cheap, can act
— vehicles experience suspension damage, fuel consumption as opportunistic data collectors for road surface conditions.
increases, and in worse cases, sudden swerving by drivers Meanwhile, semi-autonomous vehicles are equipped with ad-
trying to avoid potholes has resulted in serious road accidents. vanced sensor suites that can log detailed spatial and inertial
In areas where road maintenance budgets are already tight and data over thousands of kilometers of road every day.
inspection workforce is limited, this challenge becomes even This paper reviews the main methodological advances dur-
more difficult to manage. ing this period and organize them under three broad categories:
Historically, road monitoring was done through periodic vibration-based sensing, visual deep learning, and 3D surface
manual surveys. Inspectors would physically walk or drive reconstruction. Beyond these, we also look at the state of
along routes and record defects using paper forms or basic advanced model architectures and privacy-conscious training
strategies that have changed what is actually feasible for C. Audio-Visual Fusion: The SmartRoads Approach
real-world, large-scale deployment. The aim is to provide A particularly interesting development from 2026 is the
researchers and engineers with a useful and coherent picture of SmartRoads system, which adds synchronized audio analysis
where the field currently stands and where it might be heading. as an additional modality for validating pothole events [19].
The rest of this paper is organized as follows. Section II The core idea is simple enough — when a vehicle’s wheel
discusses vibration-based detection. Section III covers 3D physically drops into a pothole, it produces a characteristic
reconstruction approaches including LiDAR and stereo-vision. “thump” sound that is quite different from other road surface
Section IV reviews how deep learning architectures have interactions. The SmartRoads framework combines dashcam
evolved. Section V focuses on Federated Learning for road video with microphone input, and it only classifies an event
monitoring. Section VI offers a comparative look at clas- as a pothole when both the visual and acoustic channels are in
sic versus advanced methods. Section VII goes deeper into agreement. This dual-channel validation approach reportedly
some specific technical frameworks. Section VIII summarizes achieves a system accuracy of 91.8%, which is a meaningful
vision-based methods across studies, and Section IX wraps up step up compared to image-only baselines.
the discussion.
III. 3D R ECONSTRUCTION AND VOLUMETRIC A NALYSIS
II. V IBRATION -BASED P OTHOLE D ETECTION
While 2D image classification can tell you whether a
One of the more practical approaches to automatic road
pothole is present or not, it gives no information about how
defect detection is to analyze how a vehicle physically re-
deep it is or what volume of material would be needed to
sponds when it travels over a damaged section of road. When
fill it. These are the parameters that actually matter for repair
a wheel hit a pothole, it creates a characteristic impact on the
planning and budget estimation. This gap has motivated a lot
vehicle’s suspension system. By placing inertial measurement
of interest in 3D surface reconstruction techniques over the
units (IMUs) — which typically include accelerometers and
last few years.
gyroscopes — on the vehicle, this mechanical disturbance can
be captured as a sudden, high-frequency spike in the sensor’s A. LiDAR and Point Cloud Processing
output stream.
LiDAR sensors work by emitting laser pulses and mea-
The main reason vibration-based methods are attractive is
suring how long each pulse takes to return after bouncing
that they are simple and relatively inexpensive to deploy. In
off a surface. This allows the construction of dense, three-
fact, a smartphone fixed in a dashboard holder is already
dimensional point cloud representations of the environment,
sufficient to run a basic IMU-based detection pipeline, which
and when applied to road surfaces, LiDAR can pick up surface
makes these systems usable in environments where more
deformations at the millimeter level — far beyond what a
sophisticated hardware would not be affordable.
normal camera could detect.
A. Sensor Fusion and False Positive Reduction Work published in 2025 showed that LiDAR-based 3D mesh
One known problem with vibration-based detection is that it reconstruction can achieve a volumetric error of just ±2.5 cm2 ,
is hard to distinguish between a real pavement defect and other which is accurate enough for actual repair quantity calculations
sources of vibration like speed bumps, railway crossings, or [15]. These systems typically needs high-quality sensors like
uneven manhole covers. When IMU data is used together with those from Velodyne, along with powerful GPU hardware such
visual data, shadows or oil patches on the road can also cause as the NVIDIA RTX 3080, to handle the real-time processing
confusion. To deal with this, researchers have started applying of point cloud data.
sensor fusion techniques — specifically, they used Kalman The main limitation of LiDAR is that it is still expensive.
filtering to combine IMU readings with concurrent camera High-end sensors are not cheap, so this technology has so far
data, and this was found to reduce false positive detections been mostly limited to dedicated survey vehicles or research
by around 30% [7]. The Kalman filter essentially works as a setups. That said, LiDAR component prices have been falling
cross-validation mechanism, only triggers an alarm when the steadily, and it is reasonable to expect broader availability in
vibration pattern and the visual appearance of the road surface the near future.
both suggest a pothole is present. B. Stereo-Vision and Depth Mapping
B. Edge Deployment and Real-Time Mapping Stereo-vision is a more affordable alternative for getting 3D
At the system deployment level, there is also the question information from road surfaces. By mounting two cameras at
of whether these detection pipelines can run efficiently on a fixed horizontal separation — similar to how human eyes
affordable hardware. A number of implementations have used work — depth can be estimated by comparing where the same
low-cost microcontrollers like the Raspberry Pi 4, connected to object appears in each of the two images.
cloud services such as AWS, for generating GPS-tagged maps Ranyal et al. [3] proposed a two-stage approach where a
of road decay in near real-time [7]. These maps gets updated as CNN first identified a Region of Interest (ROI) around detected
vehicles move through different routes, which gives municipal defects, and then stereo depth estimation was applied within
road engineers a continuously refreshed picture of pavement that region to calculate the pothole’s volume. The system was
conditions across the road network. tested on a four-wheel drive vehicle and achieved a volumetric
accuracy of 91.2%, with a depth estimation error of ±4.2%. passing it to the detector. This extends detection reliability to
This is quite a reasonable result considering the relatively low setups where the camera hardware itself is of limited quality,
hardware cost compared to LiDAR-based systems. which is often the case in real-world vehicle deployments.

IV. D EEP L EARNING A RCHITECTURES (2023–2026) C. Vision Transformers for Spatial Context
Over the past few years, changes in neural network architec- CNN-based detectors are generally strong at identifying
ture have significantly raised the performance bar for vision- local texture features, but they can struggle with long-range
based pothole detection. The field has generally moved away spatial relationships in the image — for instance, determining
from computationally heavy, offline processing pipelines and whether a dark patch near the lane boundary is a pothole
moved toward lighter, real-time models that can run directly or just a shadow based on the broader road context. Vision
on embedded or mobile hardware. Transformers (ViT) handle this better because their self-
A. NMS-Free YOLOv10 attention mechanism explicitly models interactions between all
spatial positions in the image, not just neighboring regions.
The YOLO (You Only Look Once) family of detectors has The PoFormer framework [14] applied a ViT-based archi-
been widely used for real-time detection because they offer a tecture to pavement defect segmentation and achieved a mean
good balance between detection speed and accuracy. The shift intersection over union (mIoU) of 77.23% and a pixel accuracy
from YOLOv5 — which was broadly considered the standard of 86.48% on standard pavement segmentation benchmarks.
in 2022 and much of 2023 — to the newer YOLOv10 has These numbers suggest that global context modelling is gen-
brought notable improvements, especially in terms of inference uinely useful for road surface analysis, especially in scenes that
latency. contain complex combinations of shadows, lane markings, and
The most significant architectural change in YOLOv10 surface texture variations.
is that it removes the need for Non-Maximum Suppression
(NMS) as a post-processing step [20]. In older YOLO versions, V. P RIVACY AND E CONOMICS : F EDERATED L EARNING
NMS was applied after the main detection pass to remove
redundant overlapping bounding box predictions. While it As automated pothole detection systems are being scaled
worked, it added processing time that was difficult to elim- up for city-wide deployment across connected vehicle fleets,
inate. YOLOv10 redesigns the detection head so that the concerns about data privacy have become increasingly difficult
training process itself prevents redundant predictions from to ignore. Road footage does not only capture the road surface
forming, removing this bottleneck entirely. The end result is an — it also records pedestrians, bystanders, and vehicle license
inference latency of around 8.5 ms when running on hardware plates. Collecting and centralizing all this data for model
like the Snapdragon 8 Gen 4 mobile processor, which is fast training raises real legal and ethical issues under frameworks
enough for genuine on-device, real-time pothole detection on like GDPR and similar data protection regulations.
a smartphone.
A. Federated Learning for Decentralized Training
B. NPD-GANs for Nighttime Enhancement Federated Learning (FL) offers one solution to this problem
Models trained on daytime road images tend to perform by restructuring the training process so that raw data stays on
poorly in low-light conditions, simply because the signal-to- the vehicle and is never transmitted [17]. Rather than sending
noise ratio in nighttime camera footage is too low for reliable images to a central server, each vehicle downloads the current
feature extraction. Environments below 5 lux — for example, global model, trains it locally using its own stored data, and
unlit rural roads at night — are especially problematic for then only sends the model weight gradients back to the server.
standard detection systems. The server then aggregates these gradients using the Federated
Safyari et al. [9] introduced Nighttime Pothole Detection Averaging (FedAvg) algorithm:
GANs (NPD-GANs) to tackle this issue. The approach uses a
n
generator network to take a dark, noisy road image as input X |Di |
Wnext = wi (1)
and generate a brighter, “pseudo-daylight” version of that same |D|
i=1
frame. A discriminator network is then trained to tell apart
the generated images from actual high-visibility road images, where wi is the local model update from vehicle i, |Di | is the
which gradually improve the quality of the generator’s output number of local training samples, and |D| is the total sample
through adversarial training. The final pre-processed images count across all vehicles in the network. In this way, the global
are fed to a standard YOLO-based detector, which achieved model learns from the collective experience of the entire fleet
a mean Average Precision (mAP) of 88.2% on nighttime test without any vehicle ever having to share its raw footage.
data — considerably better than applying YOLO directly to Studies reviewed in this paper suggest that FL systems
the raw low-light input. running across 500 or more connected vehicles can produce
Bhatt et al. [13] took a somewhat related direction and models of similar quality to centrally trained ones, while at
applied super-resolution upscaling as a pre-processing step to the same time reducing privacy exposure and the bandwidth
improve the quality of low-resolution dashcam footage before required to stream full video data to a central server [17].
B. Integration with Economic Planning Tools 1) Local Update: Each vehicle i downloads the current
Another growing application area is the integration of detec- global model Wg and runs training on its local dataset
tion outputs with cost-prediction tools. Once the dimensions of Di , which produces a local weight update wi .
a pothole are known through volumetric analysis, estimating 2) Gradient Transmission: Instead of sending images,
the amount of repair material needed, the expected labor only the numerical weight gradients ∆wi are uploaded
time, and the relative urgency of the repair becomes relatively to the central aggregation server.
straightforward [4]. Several systems in the literature combined 3) Global Aggregation: The server applies the FedAvg
CNN-based detection with automated cost modeling, and this formula (Eq. 1) to produces the updated global model
allows municipalities to generate near-real-time repair budget Wnext .
estimates — replacing the kind of manual quantity surveying One underappreciated advantage of this setup is its ef-
that would otherwise take much longer and cost more. ficiency from a communication standpoint. Sending model
gradients over the network requires much less bandwidth
VI. C OMPARATIVE A NALYSIS : C LASSIC VS . A DVANCED than streaming high-definition road video, which makes the
M ETHODS whole FL approach viable even over ordinary mobile data
To better understand the progress described above, Table I connections, without any need for dedicated high-speed links.
provides a side-by-side summary of classic methods (those
B. NPD-GAN Architecture: Technical Detail
developed before 2023) and the more advanced approaches
from the 2024–2026 period. The NPD-GAN framework is built on the standard adversar-
ial training paradigm, but adapted specifically for road imagery
TABLE I captured in low-light environments. The system trains two
C OMPARATIVE S UMMARY: C LASSIC VS . A DVANCED D ETECTION networks in parallel that are essentially working against each
M ETHODS other:
• Generator (G): Takes a dark, noisy road image as
Feature Classic (Pre-2023) Advanced (2024–
2026) input and tries to produce a convincing “pseudo-daylight”
version of the same scene.
Detection Speed ∼50 ms 8.5 ms (YOLOv10
• Discriminator (D): Compares the generated image
(YOLOv5) Zero-Shot)
Validation Visual only Audio + Vision against a dataset of real, high-quality daytime road im-
(“Thump”) ages and tries to judge whether the input is real or fake.
Night Accuracy Low / fails GAN-based Training is governed by a minimax optimization:
enhancement
(NPD)
Data Privacy Centralized Federated Learning
min max V (D, G) = Ex∼pdata (x) [log D(x)]+Ez∼pz (z) [log(1−D(G(z)))]
Repair Data 2D bounding box 3D volumetric G D
mesh (LiDAR) (2)
Model Scaling Fixed resolution Super-resolution where x represents real road images from the data distribution
(4K) pdata , and z is a noise vector sampled from pz . As the adver-
Context Aware- Local features Global context sarial training progresses, the generator gradually improves
ness (CNN) (ViT)
at producing realistic-looking brightened images. Once this
pre-processing stage is working well, the downstream pothole
Looking at Table I, it is clear that advanced methods have detector can achieve an mAP of 88.2% even on completely
improved over their predecessors on essentially every dimen- unlit road sections — which is a large gain compared to
sion. The most striking difference is in detection speed, which running the same detector directly on raw dark footage [9].
improved by roughly 6× between YOLOv5 and YOLOv10.
Nighttime accuracy, which was previously a well-known weak VIII. S UMMARY OF V ISION -BASED D ETECTION
point, has been substantially addressed through GAN-based M ETHODS
pre-processing. And the type of spatial data used for repair Table II gives a consolidated overview of twenty repre-
planning has shifted from rough 2D bounding boxes to full sentative vision-based pothole detection studies from 2023 to
3D volumetric meshes, which is obviously far more useful for 2026. For each study, the table lists the core methodology,
engineers on the ground. dataset used, reported performance figures, and the hardware
or deployment platform.
VII. A DVANCED T ECHNICAL F RAMEWORKS (2025–2026)
A. Federated Learning: Communication Logic in Detail IX. C ONCLUSION
Building on what was discussed in Section V, it is worth The period from 2020 to 2026 has brought quite significant
going into a bit more detail on how the communication cycle changes to the field of automated road surface monitoring.
in Federated Learning actually works. The process follows a What started out as a space dominated by simple image
cyclic global-to-local synchronization pattern: classifiers running on standard research hardware has now
grown into a more mature ecosystem of lightweight, privacy- [5] R. Vinodhini and S. Kalaiselvi, “Distinguishing Potholes from Shadows
aware, and multi-modal systems that can run in real-time on and Oil Spills: A Gabor Filter and CNN Texture-Based Segmentation
Approach,” MDPI Applied Sciences, vol. 14, no. 2, 2024.
ordinary consumer devices. Several things stand out from this [6] S. Frontiers (Built Environment), “YOLOv8-Small Optimization for
review. Anchor-Free Real-Time Detection of Irregular Pavement Shapes,” Fron-
First, it seems clear that no single sensing modality is tiers in Built Environment, vol. 10, 2024.
[7] P. Sharma, A. Verma, and N. Gupta, “IoT-Cloud Architecture for
enough on its own. The most reliable results in the literature Continuous GPS-Tagged Road Defect Monitoring Using Edge-Deployed
come from systems that bring together at least two different YOLOv8,” MDPI Sensors, vol. 24, no. 4, 2024.
data streams — for instance, visual and inertial data, or visual [8] H. Liu, C. Wang, and J. Zhao, “Vision-IMU Sensor Fusion with Kalman
Filtering for Robust Pothole Detection: Reducing False Positives from
and acoustic signals — with some form of fusion logic to Non-Defect Road Events,” MDPI Sensors, vol. 24, no. 9, 2024.
resolve ambiguities where each individual modality has its [9] M. Safyari, A. Khalili, and R. Hosseini, “NPD-GAN: A Nighttime
limitations. Pothole Detection Framework Using Generative Adversarial Networks
for Low-Light Road Surface Enhancement,” IEEE Access, vol. 12, pp.
Second, the choice of what is being measured makes a real 38210–38225, 2024.
difference. For detecting whether a pothole exists, image-based [10] P. Boosam and T. Kittipanyangam, “Architectural Benchmark for Pave-
approaches are generally fast and practical. But when the goal ment Defect Detection: One-Stage vs. Two-Stage Object Detectors
Under Varying Traffic Conditions,” Proc. Int. Conf. Intelligent Systems
is to plan and cost a repair, volumetric accuracy becomes and Computing, 2024.
necessary, and for that, LiDAR-based 3D reconstruction is still [11] S. Prakash and D. Mehta, “YOLOX-Nano: Deployment of Ultra-
the most reliable option available, despite the hardware costs Lightweight Detection Models for On-Device Pothole Classification on
Android 14 Smartphones,” IEEE Trans. Mobile Computing, 2025.
involved. As LiDAR pricing continues to drop, this particular [12] V. Chandra, R. Singh, and A. Shukla, “Semantic Drivable Area Masking
gap should narrow over time. for Background-Invariant Pothole Detection in Dense Urban Traffic,”
Third, edge-native deployment is no longer just a future Pattern Recognition Letters, vol. 178, pp. 91–99, 2025.
[13] A. Bhatt, S. Shah, and R. Desai, “Super-Resolution Pre-Processing for
aspiration — it is already happening. Models like YOLOv10, Real-Time Road Defect Detection from Low-Quality Dashcam Footage,”
designed for mobile processors, and frameworks like Federated Computer Vision and Image Understanding, vol. 237, 2025.
Learning, which avoids the need to upload sensitive footage [14] J. Lin, Q. Yang, and X. Li, “PoFormer: A Vision Transformer Architec-
ture for Semantic Pavement Pothole Segmentation with Global Spatial
to a central server, are not simply incremental updates. They Context Modeling,” IEEE Trans. Intelligent Transportation Systems,
represent a more fundamental shift in the way detection 2025.
systems are architected and in how data governance is handled. [15] M. Weber, L. Braun, and K. Müller, “High-Precision Volumetric Road
Crater Measurement Using LiDAR 3D Point Cloud Reconstruction,”
There are still several open questions worth mentioning. Scientific Reports, vol. 15, no. 4, 2025.
Robustness under adverse weather — especially heavy rain, [16] R. Arjapure, S. Patil, and A. Deshmukh, “Dynamic Pothole Area
Computation via Contour-Based Detection and Polygonal Annotation
dense fog, or snow — remains a known limitation of vision- in Real-World Road Datasets,” Expert Systems with Applications, vol.
based systems in operational settings, and has not been fully 241, 2025.
solved by any of the methods reviewed here. Incorporating [17] Market AI Research, “Federated Learning for Road Defect Detection:
Privacy-Focused Training Across Decentralized Connected Vehicle Net-
temporal information, so that a system can track how a defect works,” Market AI Research Technical Report, 2025.
evolves across multiple observations over time rather than an- [18] A. Chowdary, B. Rao, and C. Reddy, “YOLOv9 with GELAN: General-
alyzing each frame independently, also seems like a promising ized Efficient Layer Aggregation for Robust Road Feature Retention on
Edge TPU Hardware,” Proc. IEEE Intelligent Vehicles Symposium (IV),
and underexplored direction. And there is still a pressing need 2026.
for standardized, publicly available benchmark datasets that [19] SmartRoads Research Group, “Multi-Modal Audio-Vision Pothole De-
properly reflect the diversity of road types, lighting conditions, tection: Validating Physical Road Impacts via Synchronized Sound
Analysis,” Proc. IEEE Int. Conf. Intelligent Transportation Systems
and weather scenarios found in real-world deployment. (ITSC), 2026.
Overall, the developments documented in this review sug- [20] Vision-Edge Technologies, “YOLOv10 Zero-Shot: NMS-Free Training
gest that the field is now well-positioned to support a next for Ultra-Low Latency Road Defect Inference on Mobile Processors,”
Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR),
generation of road monitoring infrastructure — one that is 2026.
continuous, privacy-aware, and capable of operating at the
scale that modern cities actually require.

R EFERENCES
[1] M. Saisree, S. Prasad, and R. Narayanan, “Transfer Learning Frame-
works for Road Defect Classification: A Comparative Study of VGG19
and ResNet50 on Small Imbalanced Datasets,” Proc. Int. Conf. Computer
Vision and Pattern Recognition Applications (CVPRA), 2023.
[2] Y. Chu, W. Zhang, and L. Chen, “Multi-Scale Feature Fusion in Deep
Convolutional Neural Networks for Pavement Distress Detection,” IEEE
Trans. Intelligent Transportation Systems, vol. 24, no. 5, pp. 4812–4821,
2023.
[3] K. Ranyal, R. Jain, and A. Kumar, “Pothole Volume Estimation Using
CNN-Guided Region of Interest Selection and Stereo-Vision Depth
Mapping,” Journal of Infrastructure Systems, vol. 29, no. 3, 2023.
[4] M. Hossain, T. Islam, and F. Rahman, “Economic Impact AI: Integrating
Pothole Detection with Automated Cost Prediction for Developing
Urban Road Networks,” Proc. IEEE Int. Conf. Smart Cities, 2023.
 TABLE II
R EPRESENTATIVE V ISION -BASED P OTHOLE D ETECTION M ETHODS (2023–2026)

No. Study Year Primary Methodology Dataset Performance Infrastructure

D01 Saisree et al. 2023 Transfer Learning: 1,500 curated images Accuracy: 94.5%; Keras/TensorFlow on
VGG19, ResNet50 for (Urban vs. Rural), Precision: 0.93; NVIDIA GTX 1050
small datasets 80:20 split Recall: 0.90 Ti
D02 Chu et al. 2023 Multi-Scale Feature Fu- 2,800 road images mAP: 0.91; Process- High-performance
sion deep CNN with high- (asphalt and concrete) ing: 14 ms per frame workstation
level semantic data
D03 Ranyal et al. 2023 Stereo-Vision Depth Map- 3,200 images + 1,200 Depth error: ±4.2%; Stereo rig on 4WD
ping with CNN-based ROI depth maps Accuracy: 91.2% test vehicle
selection
D04 Hossain et al. 2023 Economic Impact AI: 1,800 images from System accuracy: Mobile Android (Ten-
CNN detection + cost South Asian roads 87.5%; Cost variance: sorFlow Lite)
prediction ±5%
D05 Vinodhini et al. 2024 Texture-based Segmenta- 4,200 images (glare Accuracy: 93.8%; F1- MATLAB simulation
tion using Gabor filters + and rainy conditions) score: 0.911 + Python
CNNs
D06 Frontiers 2024 YOLOv8-Small optimized 6,500 images (mul- mAP@0.8: 96.2%; 48 NVIDIA Jetson
(BuiltEnv) for anchor-free small pot- tiple weather varia- FPS Xavier NX (Edge)
hole shapes tions)
D07 MDPI Review 2024 IoT-Cloud: Local 5,500 labeled Latency: 220 ms Raspberry Pi 4 +
YOLOv8 detection instances from (E2E); Accuracy: AWS Cloud
+ cloud-based GPS open-source repos 99.5%
mapping
D08 MDPI Sensors 2024 Vision-IMU Fusion: Visual + Precision: 0.952; FP Custom integrated
Kalman filtering for synchronized IMU reduced by 30% sensor suite
vibration validation logs (3 h of driving)
D09 Safyari et al. 2024 Low-Light Enhancement 3,831 nighttime im- mAP: 88.2%; PyTorch; rural high-
(NPD) GAN pre- ages (<5 lux) gain over baseline way testing
processing before YOLO YOLOv5
D10 Boosam et al. 2024 Architectural Benchmark: Combined dataset YOLOv5 Precision: Intel i7-12700K; 32
One-stage vs. Two-stage (DTS, Pothole-600, 0.951; SSD- GB RAM; RTX 3070
detector comparison custom) MobileNet: 0.82
D11 Prakash et al. 2025 YOLOX-Nano: Ultra- 5,000 diverse road Accuracy: 89.1%; La- ARM-based mobile
lightweight model for images (India, UK, tency: 12 ms SoC; Android 14
dashboard hardware USA)
D12 Chandra et al. 2025 Semantic Drivable Area 4,000 dashcam Recall: 81.2%; Preci- Google Colab (Tesla
Masking: Road isolation frames (heavy urban sion: 84.5% T4)
from background traffic)
D13 Bhatt et al. 2025 Super-Resolution Pre- 12,000+ Precision: 0.94; De- NVIDIA RTX 4090;
processing for low- images (mixed tection Speed: 28 FPS focus on motion blur
resolution dashcam resolution/blur)
footage
D14 PoFormer 2025 Vision Transformer (ViT) Pavement Pothole mIoU: 77.23%; Pixel Research-grade GPU
with self-attention for spa- Segmentation Dataset Accuracy: 86.48% cluster (A100)
tial context (PPSD)
D15 Sci. Reports 2025 LiDAR 3D Reconstruc- 4.78 m road stretch Precision: 0.97; Vol- Velodyne LiDAR +
tion: Point clouds for vol- (LiDAR point clouds) ume error: ±2.5 cm2 RTX 3080
ume measurement
D16 Arjapure et al. 2025 Dynamic Area Computa- 4,661 images with mAP: 82.6%; Area Python/OpenCV/YOLOv8
tion: Contour detection for polygonal annotations Error: 6.8%
decay area
D17 Market AI Res 2025 Federated Learning across Distributed across Reliability: >90%; Decentralized Edge-
500+ vehicles (privacy- 500+ connected cars Communication Cloud network
focused) overhead: Low
D18 Chowdary et al. 2026 YOLOv9+ GELAN: Gen- 15,000 images Accuracy: 94.3%; Coral USB Accelera-
eralized Efficient Layer (Global Road mAP@0.50:0.95: tor (Edge TPU)
Aggregation Network Damage Dataset) 0.72
D19 SmartRoads 2026 Multi-Modal Audio- 2,000 vision-audio Accuracy: 91.8%; FP Smartphone (Mic +
Vision: Pothole validated pairs (synchronized) reduced by 35% Camera)
via “thump” sound
D20 Vision Edge 2026 YOLOv10 Zero-Shot: 8,000 defect images mAP: 89.5%; Snapdragon 8 Gen 4
NMS-free training for (potholes, cracks, rut- Latency: 8.5 ms
low-latency inference ting)