Received 13 November 2024, accepted 22 December 2024, date of publication 31 December 2024, date of current version 13 January 2025.

Digital Object Identifier 10.1109/ACCESS.2024.3524676

SafeTrax: Smart Collision Prediction and Alert

System Using IoT for Sustainable Traffic Safety
MD. FAHIM BIN AMIN , ISRAT JAHAN KHAN , FARIA AKTER , ADIB AHMED,
AND MD. MOTAHARUL ISLAM , (Member, IEEE)
Department of Computer Science and Engineering, United International University, Dhaka 1212, Bangladesh
Corresponding author: Md. Motaharul Islam (motaharul@[Link])
This work was supported by the Institute for Advanced Research Publication Grant of United International University under Grant
IAR-2024-Pub-065.

ABSTRACT With the rise in traffic accidents worldwide, it has become very important to have a more
effective way to ensure road safety. To address this issue, we have made a system called SafeTrax. SafeTrax
uses deep learning to predict traffic accidents and car crashes. Additionally, it has been implemented with
the Internet of Things (IoT) to gather real-time data from sensors, and AWS Greengrass to process that data
quickly and give drivers timely warnings. Unlike old safety systems, SafeTrax can predict if an accident
might happen. It warns drivers beforehand to make roads safer for cars, people, and other animals. SafeTrax
uses video frames from camera sensors in cars to look for dangers. It then uses a special deep-learning
program to figure out if a crash might happen. One important part of SafeTrax is the integration of AWS IoT
Greengrass. It helps the system work faster by using cloud services on devices like the Raspberry Pi. This
means the system can quickly check the data and send warnings in time. SafeTrax can ensure safer roads,
and it can help save lives.

INDEX TERMS Traffic, collision, SafeTrax, AWS IoT Greengrass, deep learning, IoT, pedestrian.

I. INTRODUCTION The challenging endeavor is to build a system that can

Metropolitan regions that are rapidly broadening, gridlocked predict accurately and prevent catastrophes. The dynamic
roads, and vehicular crashes have become significant issues and unpredictable nature of road conditions can quickly
for modern-day society. The frequency of catastrophes has turn routine driving scenarios into hazardous situations [5].
spiked as more cars are on the road, which may result The economic challenges and modern technologies show the
in significant consequences like loss of life, and monetary obstacles in the way of developing safer roadways [6], [7].
difficulties for families [36]. Traffic accident prevention has The dynamic circumstances of traffic and external variables
been an essential concern due to these worrisome statistics, such as weather aggravate these problems, emphasizing the
prompting cutting-edge technologies that may reduce the risk need for innovative, cost-effective solutions [8], [9], [10],
of crashes [1], [2], [35]. Authorities as well as companies are [11].
implementing progressively stricter laws regarding traffic, In recent years, so many better models with highly
but there is still an urgent requirement for solutions that effective factors and frameworks have been developed
may predict and avoid collisions from happening before by a couple of researchers to resolve such challenging
they strike. This emphasizes the significance of establishing scenarios and to prevent crash scenarios on roadways [37].
cutting-edge technology that improves drivers’ awareness Traditional methods often come up short of maintaining
of their surroundings in real time by warning them about the level of complexity since they rely upon human atten-
potential dangers beforehand [3], [4]. tion along with straightforward resources. Several methods
like the Chain of Road Traffic Incident (CRTI) frame-
work, Deep Spatio-Temporal Graph Convolutional Networks
The associate editor coordinating the review of this manuscript and (DSTGCN), and Hybrid Learning Networks have been
approving it for publication was Tariq Umer . introduced for this purpose [12], [13], [14], [15]. These
2024 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
VOLUME 13, 2025 For more information, see [Link] 6667
 M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

methods use machine learning and artificial intelligence highly adaptable and affordable model for all kinds of
(AI) to improve how well accidents can be predicted. vehicles.
These systems are capable of predicting potential collisions • We have contributed to the democratization of road
and giving early warnings to drivers by analyzing data safety by developing a system that can operate across
from vehicle movements, weather conditions, and traffic both urban and rural contexts, greatly improving acci-
flow [16], [17]. Despite being of great progress, some of dent prediction accuracy and providing timely alerts to
these approaches are still not fully efficient as they lack drivers, ensuring broad and effective coverage.
the scenarios of testing these in real-world situations or do • Our core Deep learning architectures have shown strong
not incorporate IoT technologies that allow for real-time deep learning performance. ResNet50 has achieved
transmission [18]. 93.03% accuracy and a 95.57% F1-score, while
Recognizing these challenges, our research offers a MobileNetV2 has slightly outperformed with 93.3%
proactive and real-time approach to predict whether any accuracy and a 96.08% F1-score. VGG-16, though
possible collision can occur or not. In contrast to the slightly lower, has still performed well with 90.75%
majority of earlier research, which concentrated on collision accuracy and a 93.72% F1-score.
detection by identifying accidents after they occur. To greatly The rest of this paper is organized as follows. In Section II,
improve road safety, we have developed a state-of-the-art we introduce related work about some previous research
collision prediction system in our research that makes use closely aligned with our field. Section III proposes infor-
of deep learning models and IoT-based communication. The mation about the dataset and methodologies including
transition from detection to prevention has proven to be MobileNetV2, ResNet50, VGG16, etc., with their model
pivotal, as it enables prompt notifications and proactive architectures and algorithms. Experiments in Section IV
measures, providing drivers with a vital window of time to verify the feasibility and performance of the proposed
avert accidents completely. method.
Affordability while being a practical solution, has been
the main focus of our work. To contrast a complete safety
system, we have created an effective pipeline that smoothly II. RELATED WORKS
integrates ambient data, real-time from onboard sensors, and We have studied some previous research closely aligned with
deep learning models. Vehicles can enhance their object our field. We have presented our in-depth analysis of them in
detection capabilities and collision prediction using SafeTrax this section. However, this section is divided into two parts:
by utilizing real-time sensor data, leading to improved literature review and gap analysis.
accuracy and responsiveness of the prediction system. Our
work not only pushes the boundaries of existing road safety A. LITERATURE REVIEW
technologies but also demonstrates how deep learning and Recent advancements in traffic accident prediction have
IoT systems can be effectively utilized. leveraged various computational models and frameworks
The significance of this research lies in its potential to to enhance accuracy and efficiency in complex traffic
make far-reaching changes in road safety. By incorporating scenarios. Xiong et al. [19] have introduced the innovative
such a system that predicts collisions and alerts drivers Chain of Road Traffic Incident (CRTI) framework for
in advance, we have offered a practical solution that can improved accident prediction, leveraging vehicle movement
be implemented in a wide range of scenarios with less data through a two-pronged approach: an initial separation
economic challenges. Our work has pushed the boundaries of vehicular lane behavior via a Support Vector Machine
of existing collision detection technologies and also has (SVM), followed by differentiation between accident and
provided a blueprint for how deep-learning models and IoT- non-accident events using Gaussian-mixture-based Hidden
based communication systems can work together in harmony Markov Models (HMM). For accident prevention under
to create safer roadways. Addressing a critical need in today’s complex traffic scenarios, their findings validated by PreScan
increasingly congested traffic environments, this innovative platform simulations, have underscored the CRTI frame-
approach aims to mitigate traffic accidents and ultimately work’s potential to enhance early warning systems.
save lives. Malawade et al. [20] have critiqued past collision pre-
The major contributions to our work have been mentioned diction methods for autonomous vehicles, highlighting the
below: flaws in early statistical metrics and model-based approaches
• We have proposed a forward-thinking solution that due to unrealistic assumptions and high false-positive rates,
integrates IoT and Deep Learning by equipping vehicles respectively. To more accurately predict collision, they have
with sensors and dashcams, and it creates a network proposed scene-graph-based methods by integrating Multi-
system that continuously feeds real-time data into Relational Graph Convolutional Networks (MR-GCN) with
a central hub where sophisticated algorithms predict LSTM for spatial and temporal analysis, and collisions by
potential collisions with high accuracy. understanding dynamic inter-object relationships. Although
• We have analyzed a predictive algorithmic process to specific results have not been detailed, this approach
predict potential accidents or collisions by using our has shown improved prediction accuracy and timeliness.

6668 VOLUME 13, 2025

M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

Venkateswaran et al. [21] have explored how deep learning, by capturing spatiotemporal patterns. The results have
particularly Convolutional Neural Networks (CNNs), has indicated the superior performance of the DSHN model
revolutionized vehicle detection, tracking, and distance over traditional and machine learning models, evidenced
estimation for forward collision warning systems. They have by improved precision, recall, F1-score, and AUC metrics.
showcased methods like R-CNN, Fast R-CNN, and YOLO Thapa et al. [29] have improved duration-based models to
for detection and have employed correlation filters with predict crash likelihoods and severities by integrating both
deep learning for the improved tracking of dynamic objects. static and dynamic factors like traffic and weather conditions.
Planar homography and Inverse Perspective Mapping (IPM) Using data from Tennessee, they developed a validated model
techniques have been discussed for distance estimation, with that has shown effective prediction with a balanced dataset,
IPM’s accuracy having been boosted through road boundary highlighting the importance of certain variables for accuracy.
detection and automatic calibration. Zhang et al. [22] have In the study by Rathore et al., [30], a smart traffic control
reviewed various motion prediction models, distinguishing model has been presented, utilizing vehicular cameras and
them as physics-based, maneuver-based, and interaction- fog computing for real-time traffic violation monitoring.
aware models. Their methodology has focused on evaluating The approach has employed GPU-based fog devices with
models like Constant Velocity and Constant Acceleration Apache Spark for processing, leveraging techniques like SSD
for real-time applications, and maneuver-based models using for vehicle detection and the Sobel-Feldman edge detection
techniques such as SVMs and HMMs, while also considering for lane identification to detect violations such as illegal
interaction impacts between vehicles. They have highlighted U-turns and driving on central dividers. By implementing
the improvement in prediction accuracy through the use of it on a GTX750-Ti GPU machine, the system has shown
Interactive Multiple Models (IMM) which blend physics high accuracy and efficiency in urban traffic management.
and maneuver-based approaches for dynamic scenarios. The Pramanik et al. [31] have designed a video surveillance
research has confirmed the effectiveness of these models in system to enhance road safety, incorporating five algorithms
real-world traffic scenarios, particularly in assessing risks for (SVD, OTD, OD, IPD, and WDLPD) for detecting traffic
lane changes and rear-end collisions. pre-events like speed violations and illegal parking. Their
Gus et al. [24] have explored Advanced Driver Assistance methodology has demonstrated effectiveness by analyzing
Systems (ADAS) technology, focusing on forward-collision 132 real-time traffic videos from India, outperforming current
warnings, and have employed the Gaussian Mixture Model methods in pre-event detection. Islam et al. [32] have
(GMM) for time-to-collision estimates and used LSTM and developed a smart device that can detect motorcycle accidents
Deep Belief Network (DBN) models to predict collision and facilitate immediate assistance for injured individuals.
risks. Their methodology has been tested on real-world The device aims to minimize casualties by quickly sending
urban freeway data, demonstrating the predictive accuracy alert messages to hospitals, police stations, and registered
of these models. Dong et al. [25] have developed a family members after an accident. This initiative addresses
model blending unsupervised learning to uncover complex the critical gap in timely medical response to motorcycle
interactions in traffic data, and supervised learning through crashes in Bangladesh, where road accidents and fatalities are
negative binomial regression, effectively enhancing crash rising.
prediction accuracy. When applied to Tennessee data, their
dual approach has significantly outperformed traditional B. BENCHMARK OF PREVIOUS RESEARCH WORKS
models. Using sensors to detect accidents and report their
This section presents a benchmark of related research works
locations via GPS. Sahoo et al. [26] have developed an
and compares how the previous studies implemented their
IoT-based automated system for rapid and accurate traffic
proposed methodologies and how well they performed.
crash reporting. Significantly reducing response times and
By doing this, we have decided what has improved in
enhancing the effectiveness of emergency services, their
road safety and what still needs more research. This
approach has integrated low-cost technologies to improve
comparison has clarified various methodologies for us and
vehicle safety. Yu et al. [27] have introduced the Deep
illuminated how our research could contribute to enhanc-
Spatio-Temporal Graph Convolutional Network (DSTGCN)
ing traffic safety. Table 1 shows results and key points
for traffic accident prediction, employing data from various
of previous studies on road safety and traffic accident
sources such as traffic records and meteorological conditions.
detection.
In this model, spatial and temporal data analysis have
been combined with external data integration to demonstrate
notable advancements in traffic safety and management by C. GAP ANALYSIS
outperforming existing models. In reviewing past studies on road safety, and the detection
Kashifi et al. [28] have introduced a Deep Hybrid and forecasting of traffic accidents, it has been found that
Learning Framework to forecast short-term traffic crashes there is a significant reliance on older techniques and
through a Deep Spatiotemporal Hybrid Network (DSHN) costly setups. This review emphasizes the critical need for
that integrates CNNs and LSTMs. This approach has improved and more readily available approaches that utilize
processed varied traffic data, increasing prediction accuracy the latest technology to enhance the accuracy of accident

VOLUME 13, 2025 6669

 M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

TABLE 1. Benchmark of literature review.

prediction systems and improve road safety. Table 2 offers object is detected, the vehicle’s sonar sensors are activated
a comprehensive overview of the results and limitations of to measure the distance between the vehicle and the object.
past studies, indicating the areas that still have the scope for When the sonar sensors find a possible collision, they check
further research and development. how close nearby objects are. If the object gets too close to
the car, the IoT system alerts the driver with an alarm. These
III. PROPOSED SYSTEM sensors also help with spotting obstacles. By using real-
In this section, we have provided the dataset information, time data from the sensors, our system helps vehicles notice
detailed methodology, architecture details, and algorithm of nearby collision threats. These sensors work with predefined
our work. This section is divided into seven subsections: safety distances; if the distance falls below a safe margin,
System Overview, Dataset, Model Architecture, IoT Archi- the system classifies it as a high-risk situation and generates
tecture, IoT Implementation Process, Algorithm, and AWS an immediate alert to warn the driver. The combination of
Greengrass in SafeTrax. Deep Learning, IoT, and AWS Greengrass enables accurate
collision predictions, reduces false positives, and enhances
A. SYSTEM OVERVIEW real-time decision-making, adapting to changing traffic con-
ditions through a continuous feedback loop. This proactive
Our system integrates advanced IoT devices, deep learning,
approach aims to significantly reduce accidents by providing
and AWS Greengrass for a real-time traffic safety solution.
early warnings to drivers and facilitating quick evasive
The car has IoT sensors and dash cameras that gather
actions. Figure 2 illustrates the simple workflow diagram of
important information about its surroundings, like how close
SafeTrax.
it is to other vehicles. Raspberry Pi collects real-time data
from all the connected sensors and cameras. At the central
hub, trained deep-learning models (MobileNetV2, VGG-16, B. DATASET
and ResNet50) are used. These models check if an accident The BDD100K dataset, which contains 100,000 images of
might happen using data from the cameras. The camera car collisions, has been taken as the main dataset for our
modules continuously monitor the surroundings by providing study. We have focused on a subset of 10,000 images for
real-time video input for the deep-learning models. They binary classification which have been annotated manually by
analyze the camera data to identify objects like cars, people, us. We categorize them into two classes: ‘Collision’ (label
and other structures. Deep learning models are set up on ‘y’) and ‘No Collision’ (label ‘n’). This dataset has been
the Raspberry Pi using AWS Greengrass, which lets the Pi divided into training (70%), validation (10%), and testing
run these models right on the device for quick analysis and (20%) subsets. It has allowed for robust model training
sending alerts. AWS Greengrass also helps the Raspberry Pi and evaluation. The careful annotation and organization
connect smoothly to the cloud. It allows real-time decision- aim to support effective model development in traffic
making on the device and more complex tasks to be done accident detection and analysis. Our customized dataset can
in the cloud. Once the deep learning models detect objects, be accessed on Kaggle. Table 3 represents the necessary
the system moves to the next stage of analysis. If a potential information for the BDD100K dataset.

6670 VOLUME 13, 2025

M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

TABLE 2. Gap analysis table.

1) DATASET PREPROCESSING rotation angles ranging from 0 to 75 degrees. It introduces

Preprocessing our custom dataset is crucial for preparing the angular variation and helps the model handle different
data for model training and testing. The applied preprocessing orientations. Additionally, zooming techniques have been
steps are given below: applied to create a clearer view of distant objects. It allows
the model to better recognize and learn from objects at various
2) DATA CLEANING distances.
Before starting the preprocessing steps, it is important to
clean the dataset. It involves removing corrupted or damaged b: RESIZING
images, as well as filtering out images that are not relevant All images have been resized to an exact resolution of 256 ×
to our system. These images could cause confusion during 256 pixels. This uniform size ensures that all inputs to the
the training of deep learning models. These actions have model are the same. It makes training and processing more
enhanced the quality of our dataset and ensured more efficient.
effective training.
c: NORMALIZATION
a: DATA AUGMENTATION After resizing, the pixel values of the images have been
To make the model more reliable, we have applied data normalized to a range between 0 and 1. It has been
augmentation techniques. We have applied horizontal flip- done by dividing each pixel value by 255. This step has
ping to some images to introduce variability in object improved the training process and helped the model run more
orientation. We have also utilized random rotation with effectively.

VOLUME 13, 2025 6671

 M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

3) LABELING PROCESS PIPELINE FOR PRE-COLLISION b: THRESHOLDING FOR PRE-COLLISION DETECTION AND
DETECTION AUTOMATIC LABELING
We have employed an automated and systematic process To determine whether an image represents a potential pre-
for labeling images for pre-collision detection. This has collision scenario, the estimated distance has been compared
been done by measuring the distance of objects from to predefined thresholds. For vehicles, the threshold has been
the camera modules and assigning labels to each image. set within a range of 12 to 18 meters, while for humans, it has
This fully automated method has ensured accurate label- been set within a range of 8 to 12 meters. If the estimated
ing of the dataset and eliminated the need for manual distance is below the set threshold, the image has been labeled
labeling. as a pre-collision scenario (‘y’); if it is not, it has been labeled
as a non-pre-collision scenario (‘n’).
a: CAMERA CALIBRATION AND OBJECT PROPERTIES This range-based approach allows the system to account
The first step involves determining the properties of the for varying object sizes, enhancing the robustness of the
camera modules and the real-world dimensions of the objects. labeling process. Figure 1 illustrates the overview of how we
The focal length of the camera modules has been estimated have made annotations to our custom dataset.
based on the following properties:
• Field of View (FOV): 120Â◦ .
• Image Resolution: 1080p (1920 × 1080 pixels).
• Focal Length (f): The focal length has been calculated
using the formula:

Image Width
f =  
2 × tan FOV
2

For estimating object size, we have defined ranges to

account for the differences in the dimensions of vehicles and
humans:
• Car Width Range: 1.6 to 2.2 meters.
• Human Height Range: 1.5 to 2.0 meters.
Bounding Box Detection and Object Classification
Using a pre-trained object detection model (YOLOv8),
bounding boxes are generated around vehicles and humans
for each image. Along with the bounding boxes, the object
types (e.g., car or human) are classified. The size of the object
in pixels is extracted from the bounding box:
• Vehicles: The width of the bounding box is used.
• Humans: The height of the bounding box is used.
Distance Estimation Using Pinhole Camera Model The FIGURE 1. Pipeline diagram of the dataset annotation process.
distance to each object is estimated using the pinhole camera
model:
C. MODEL ARCHITECTURE
Complete Object Size × f Our model architecture has been designed as a single pipeline
Distance =
Object Size in Pixels focusing on deep learning-based accident prediction and
IoT-based alert system integration. Therefore, this subsection
where:
has been divided into four parts: ResNet50 Architecture,
• Complete Object Size refers to the object’s size based VGG16 Architecture, and MobileNetV2 architecture.
on its classification (vehicle or human), chosen within
the defined range (1.6m - 2.2m for vehicles, 1.5m - 2.0m 1) RESNET50 ARCHITECTURE
for humans). We have used the ResNet50 model architecture, which has
• Object Size in Pixels is the size of the object in the been trained on the ImageNet dataset and accepts an input
image (width for vehicles, height for humans) which has size of 224 × 224 × 3. The model begins with a 7 ×
been determined from bounding box size. 7 convolution with 64 filters and a stride of 2, followed
• f is the focal length calculated earlier. by a max-pooling layer. ResNet50 includes residual blocks
This approach ensures the system can handle objects of with increasing filter counts across four stages: 3 blocks with
varying sizes. 64 filters, 4 blocks with 128 filters, 6 blocks with 256 filters,

6672 VOLUME 13, 2025

M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

TABLE 3. Overview of BDD100K dataset. or sigmoid activation. Figure 5 represents simple VGG-16
model architecture.

3) MOBILENETV2 ARCHITECTURE
The model architecture uses pre-trained MobileNetV2 as
the backbone feature extractor. It takes input images sized
256 × 256 pixels with 3 color channels. MobileNetV2
outputs a feature map of 8 × 8x1280. To enable binary
classification, a Global Average Pooling layer reduces the
feature map to a 1280-dimensional vector. This is followed
by a Dense layer with 128 units and ReLU activation to
capture non-linear relationships. A Dropout layer with a rate
of 0.5 is also included to prevent overfitting. Finally, a Dense
layer with a single unit and Sigmoid activation produces a
probability score for binary classification. Figure 3 represents
simple MobileNetV2 model architecture and table 4 provides
breakdowns of the ResNet50, VGG-16, and MobileNetV2
architecture.

TABLE 4. Architecture of ResNet50, VGG-16, and MobileNetV2 for binary

classification.

FIGURE 2. Workflow diagram of the research work.

and 3 blocks with 512 filters. A global average pooling layer

generates a 2048-dimensional feature vector, and a custom
linear layer replaces the fully connected layer to match
the number of classes in the dataset. In feature extraction
mode, the convolutional layers are frozen, which allows
only the weights in the fully connected layer to be updated
during training. Figure 4 represents a simple ResNet50 model
architecture.

2) VGG-16 ARCHITECTURE
The pre-trained VGG-16 architecture has 13 convolutional D. IoT ARCHITECTURE
layers and 3 fully connected layers that are designed for To dive into the IoT component, we need to discuss all
image classification. It accepts 224 × 224 × 3 RGB images the sensors, modules, and hardware required to create the
and includes five max pooling layers: 64, 128, 256, and prototype. We have used a Raspberry Pi, an ultrasonic
512 filters in the blocks, all using 3 × 3 kernels with ReLU sonar sensor, a Raspberry Pi camera module, LEDs, buzzers,
activations. The output is flattened and processed through and wires. This hardware setup provides a robust platform
three fully connected layers, the first two having 4096 units for implementing the pre-collision alert system, combining
and Dropout for regularization. The final layer is adapted visual input, distance sensing, processing power, and alert
for binary classification. It produces 2 units with a softmax mechanisms. The Raspberry Pi’s capabilities allow for

VOLUME 13, 2025 6673

 M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

FIGURE 3. Model architecture diagram for MobileNetV2 [33].

FIGURE 4. Model architecture diagram for ResNet50 [34].

FIGURE 5. Model architecture diagram for VGG-16 [34].

running complex deep learning models, while the additional 2) CAMERA MODULE
sensors and output devices create a comprehensive safety The Raspberry Pi Camera Module connects to the CSI port
system. of the Raspberry Pi. We used the proper CLI commands to
enable the camera. We have ensured good lighting for better
E. IoT IMPLEMENTATION PROCESS
image capture and regularly tested the camera to confirm it
works properly.
1) EDGE DEVICE (RASPBERRY PI 4)
The Raspberry Pi 4 is the main processing unit in our
system. It runs AWS IoT Greengrass Core software for local 3) SONAR SENSOR (HC-SR04)
data processing and inference. Our pre-trained deep learning The sonar sensor connections use 3.3V logic levels. Since we
models on the Raspberry Pi enable real-time collision risk have used a 5V sensor, a voltage divider has been needed
assessments using input from connected sensors. for the ECHO pin. We have also ensured that the sensor

6674 VOLUME 13, 2025

M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

enhancement in urban environments. Algorithm 1 is the

detailed algorithm outlining our entire research process.
The algorithm is well-structured. It has defined its
inputs, outputs, and purpose clearly through the procedure
named ‘‘Car_Accident_Prediction’’. The main loop operates
continuously while the system is active. In the loop,
it encompasses key steps such as data retrieval, model
prediction, object detection, and alert generation. In the
algorithm, the variable named ‘‘distScore’’ is a metric that
evaluates the distance of detected objects (such as vehicles or
pedestrians) based on their predicted properties. Additionally,
the algorithm uses a generalized distance calculation through
FIGURE 6. Overview diagram of IoT-based alert system. ‘‘ComputeDistance,’’ which adds flexibility. The use of
‘‘current alerts’’ for managing notifications is effective,
while the ‘‘TransmitAlerts’’ function correctly informs both
is securely mounted and correctly positioned for accurate nearby vehicles and the driver, which is crucial for real-
distance measurements. Additionally, we have tested the time operations. To further improve the algorithm, adding
sensor in different environments to improve its reliability. error handling, performance metrics, and clear termination
conditions could enhance its reliability and scalability.
4) BUZZER
We have used an active buzzer, so the connections are G. AWS IoT GREENGRASS IN SAFETRAX
straightforward. The buzzer is connected to GPIO 22, and we AWS Greengrass enables SafeTrax to work independently
have ensured it is rated for the voltage used and can handle and make faster decisions at the edge, even in low-bandwidth
the current from the Raspberry Pi GPIO pin. We have also conditions. It allows data processing, device management,
tested it at different frequencies to achieve the desired sound. and machine learning locally. At the same time, devices
remain connected to the cloud for updates and long-term data
5) LED storage. This helps centralize the entire system and enhance
We have used a 220 ohm resistor with our LED. The resistor its effectiveness. AWS IoT Greengrass enhances SafeTrax in
has been placed between the GPIO pin and the LED’s anode this way:
to prevent excessive current flow. We have ensured that 1. Data Collection: SafeTrax integrated vehicle has a cam-
the LED polarity is correct. Table 5 summarizes all the era module for image data and a sonar sensor for measuring
pin number details for the Raspberry Pi and its connected object distance. AWS IoT Greengrass preprocesses this image
modules. data locally by filtering out noise and preparing it for analysis.
2. Local Edge Processing: Greengrass can enable Rasp-
TABLE 5. Raspberry Pi 4 connections to peripherals. berry Pi to perform local edge processing by running deep
learning models directly on the device. The system uses sonar
data to detect objects that are dangerously close and images to
identify obstacles. By executing these models locally, it can
provide real-time feedback.
3. Cloud-Based Deep Learning: Preprocessed data can
be uploaded to the cloud for advanced analysis using deep
learning models on AWS. These models predict potential
collision risks by analyzing patterns from multiple vehicles.
This transfers heavy computational tasks to the cloud, so edge
Figure 3, figure 4, and figure 5 illustrates a detailed devices can concentrate on real-time processing.
diagram of our MobileNetV2, ResNet50, and VGG-16 model 4. Real-Time Alerts: If there has been a lack of internet
architecture respectively. These figures also showcase their connectivity, the system has to rely on the edge processing
various components and interactions, while Figure 7 presents performed by Greengrass to continue providing essential
the IoT architecture diagram. alerts based on the locally collected data.

F. ALGORITHM 1) IMPLEMENTATION PROCESS OF AWS IoT GREENGRASS

Our research workflow integrates three deep-learning models We have installed the AWS IoT Greengrass Core software on
to predict whether a collision will occur or not based on our Raspberry Pi and set up a Greengrass group in the AWS
image features. This approach has leveraged IoT frameworks IoT console. Key components of our system include a Python
for real-time data aggregation and utilized smart traffic script as the core of the application and deep learning models
control systems for automated alert generation and safety for handling machine learning tasks. We have configured

VOLUME 13, 2025 6675

 M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

FIGURE 7. Architectural diagram of IoT portion.

a timestamp. This information is sent to AWS IoT Core to

ensure consistent availability for further analysis. We have
implemented error reporting to AWS IoT Core to facilitate
remote monitoring and diagnostics. This system allows us
to track issues in real time and enables prompt responses
to errors. By sending error logs and relevant information to
the cloud, we can efficiently analyze and address problems.
Additionally, we have utilized Greengrass’s security features
to ensure our devices are secure when communicating with
AWS IoT Core. We have implemented device authentication
to verify the identity of each device and employed encrypted
communication to protect data sent over the network. These
methods help prevent unauthorized access and protect against
data breaches. Figure 8 represents the workflow diagram of
AWS GreenGrass. Figure 10 represents the diagram for the
working procedure of the key functions.
FIGURE 8. Workflow diagram of AWS GreenGrass.

FIGURE 9. Diagram design of a SafeTrax integrated car.

these models as Greengrass components for local inference

within the main application. This setup enables us to assess FIGURE 10. Diagram for working procedure of key functions.
collision risks in real-time. We have also set up local storage
to save data on collision risks and sensor readings. To keep
everything updated, we implemented a process to sync this IV. PERFORMANCE EVALUATION
data and periodically upload it to AWS IoT Core. Each data This section outlines the evaluation metrics used in our
message includes the measured distance, collision risk, and research and provides a detailed assessment of SafeTrax’s

6676 VOLUME 13, 2025

M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

Algorithm 1 Car Accident Prediction and Alert System provides a summary of the equipment and programs used
Input: Real-time sensor data stream S, Pre-trained model in our study. The most important part of our system is the
M , Distance threshold T NVIDIA GeForce RTX 3060 GPU with 12 GB of memory.
Output: Real-time alerts A Additionally, we have used TensorFlow with CUDA 12.6 to
speed up deep learning training. Python 3.12 has been the
procedure Car_Accident_Prediction main programming language in our code. We also used
Initialization: libraries like Scikit-learn, Keras, and Numpy to build models
Initialize empty alert set A = {} and manage data. The system runs on Linux, which provides
a stable environment for running deep learning tasks and
Main Loop: monitoring GPU performance.
while System is active do
X ← GetNextDataFrame(S) TABLE 6. System environment setup for deep learning experiments.
Ypred ← M .Predict(Preprocess(X ))
objects ← [o for o ∈ DetectObjects(Ypred )
if distScore greater than THRESHOLD]
current_alerts ← []
for each pair (i, j) in detected_objects do
d ← ComputeDistance([Link], [Link])
if d < T then
current_alerts.append(GenerateAlert(i, j, d))
end if
end for
TransmitAlerts(current_alerts) B. EVALUATION METRICS
end while To evaluate the performance of our deep learning models,
we have used several popular evaluation metrics that are
Helper Functions: widely used for evaluation purposes in machine learning and
Preprocess(X ): deep learning. The metrics we have used include accuracy,
return ApplyDataTransformations(X ) precision, recall, and F1 score. A brief overview of these
metrics is provided below:
GenerateAlert(object1, object2, distance): • Accuracy score: The accuracy score is one of the
return {‘‘object1’’: object1, ‘‘object2’’: object2, most commonly used metrics in the fields of machine
‘‘distance’’: distance, learning and deep learning. It represents the proportion
‘‘timestamp’’: CurrentTime()} of correctly predicted instances (both true positives
and true negatives) out of the total instances. These
TransmitAlerts(alerts): total instances include true positives, true negatives,
for each alert in alerts do false positives, and false negatives. This metric provides
BroadcastToNearbyVehicles(alert) a comprehensive assessment of the model’s overall
NotifyDriver(alert) performance.
end for
Number of Correct Predictions
end procedure Accuracy =
Total Number of Predictions
• Precision: Precision value evaluates the accuracy of
the model’s positive predictions. The positive prediction
performance in deep learning and IoT, as well as AWS means the ratio of true positive predictions to the total
Greengrass’s effectiveness in our system. The section predicted positives (true positives and false positives).
has been divided into six parts: System Environment This metric provides an overview of how reliable the
Setup for Deep Learning Experiments, Evaluation Metrics, model is when classifying instances as positive.
Performance Evaluation of Deep Learning Architectures,
True Positives
IOT Performance Analysis, AWS Performance Analysis, Precision =
and SafeTrax Integrated Vehicle Safety System: Prototype True Positives + False Positives
Design. • Recall: Recall can be denoted as the ratio of true positive
predictions to the total actual positives. Generally, it only
measures the model’s ability to identify all relevant
A. SYSTEM ENVIRONMENT SETUP FOR DEEP LEARNING
instances.
EXPERIMENTS
Table 6 shows the hardware and software we have used True Positives
Recall =
to train and test our deep learning models. The table True Positives + False Negatives

VOLUME 13, 2025 6677

 M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

TABLE 7. Performance of models on the test set.

TABLE 8. Performance of models on the validation set.

• F1 score: The F1 score is a metric that combines 91.44% (validation). Also, the F1-score is 96.08% (test).
precision and recall and provides a performance result. Figure 20 and 21 represent the accuracy and loss curves
It indicates how good a model is at predicting both for the training and validation processes generated
positive and negative cases. It is especially useful when during the training of the MobileNetV2 model.
the classes are imbalanced, • VGG-16: VGG-16 has shown lower overall perfor-
Precision × Recall mance. It has achieved an accuracy of 90.75% (test)
F1 Score = 2 × and 92.1% (validation). Its precision was 95.78% (test)
Precision + Recall
and 88.1% (validation), with recall values of 91.74%
C. PERFORMANCE EVALUATION OF DEEP LEARNING (test) and 88.4% (validation). And its F1-score is
ARCHITECTURES 93.72% (test). Figure 22 and 23 depict the accuracy
The performance of three deep learning architectures con- and loss curves for the training and validation processes
sisting of ResNet50, VGG-16, and MobileNetV2 has been generated during the training of the VGG-16 model.
evaluated. The results show that the models have effectively
learned the underlying patterns and achieved satisfactory 3) CLASSIFICATION PERFORMANCE METRICS
accuracy and loss reduction. In addition to accuracy, key metrics such as precision,
recall, and F1-score have been calculated to evaluate
1) TRAINING DYNAMICS classification performance. These metrics demonstrate
During training, the models have improved accuracy with the model’s effectiveness across all classes. Tables 7
each epoch, effectively learning and optimizing their weights. and 8 summarize results from the training and evaluation
Validation accuracy has remained stable with minimal phases for the three Deep Learning models. Figure 24
overfitting. It indicates good performance on unseen data. depicts the confusion matrix for the ResNet50 model’s
Both training and validation losses have decreased. How- test performance, while Figures 25 and 26 show the
ever, MobileNetV2 has exhibited significant fluctuations in confusion matrices for the MobileNetV2 and VGG-16
validation accuracy and loss and formed a zigzag pattern. models, respectively.
This suggests inconsistent training, possibly due to an
incorrect learning rate, overfitting, or insufficient training D. IoT PERFORMANCE ANALYSIS
data. In this section, we have presented the result analysis
produced by our IoT system and found the outputs
2) DETAILED MODEL PERFORMANCE METRICS in our Raspberry Pi OS CLI. Figures 11, 12, and 13
We have evaluated the performance of the modified archi- represent the output logs from three different states: the
tectures using a range of metrics, including accuracy, loss, no-collision state, the collision state, and the specific
precision, recall, and F1-score: timestamp with the system closing.
• ResNet50: Resnet50 has achieved an impressive accu-
racy of 93.03% on the test set and 92.6% on the 1) NO COLLISION STATE
validation set. It has demonstrated high precision Each log entry has included a timestamp for track-
(93.76% test, 90.11% validation) and recall (97.45% ing events. ‘‘INFO’’ entries have detailed operations,
test, 91.3% validation), resulting in an F1-score of while ‘‘DEBUG’’ entries have provided troubleshooting
95.57% for the test set. Figure 18 and 19 represent the information for specific camera frames. The collision
accuracy and loss curves for the training and validation probability has ranged from 0 to 1 which indicates
processes generated during the training of the ResNet50 risk levels. In Figure 11, a Collision Probability of
model. 0.023 in Frame 1 suggests a low collision likelihood,
• MobileNetV2: MobileNetV2 have gained an accuracy with a Distance of 352.70 cm. It indicates a safe
of 93.3% (test) and 90.1% (validation). It exhibited separation. An Alert Status of ‘‘OFF’’ means no col-
precision values of 94.79% (test) and 90.21% (val- lision threat has been detected, allowing for continued
idation) alongside recall rates of 97.40% (test) and monitoring.

6678 VOLUME 13, 2025

M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

1) LATENCY AND RESPONSE TIME

Latency and response time are critical for real-time
systems. In SafeTrax, IoT devices have sent data to the
Greengrass Core for local processing, which forwards
results to the AWS IoT Core. Key latency metrics like
Round-Trip Time, Sensor Input-to-Action Time, and
FIGURE 11. CLI output when no collision is detected. Cloud Processing Time have been logged. Greengrass
Core has retrieved and filtered these logs for the last
10 minutes and displayed them in the CLI. The process
2) COLLISION STATE for how AWS Greengrass retrieves these data using a
The log entry in Figure 12 has shown that, during Frame CLI command from the user-end has been presented
845, the collision probability is 0.892, with a distance in figure 14. The output after running this command is
of 48.30 cm to the nearest object. The alert status is given in table 9.
ON, activating the buzzer and LED indicators. This According to the table 9, An average RTT of 52 ms
‘‘DEBUG’’ message has confirmed that the frame data, ensures efficient communication for real-time applica-
collision probability, and alert status have been sent to tions. The 35 ms sensor-to-action latency enables timely
AWS IoT Core for further analysis. alerts, while the 20 ms cloud processing time indicates
fast data handling.

FIGURE 12. CLI output when collision is detected.

3) THE SPECIFIC TIMESTAMP WITH THE SYSTEM

CLOSING
The log message in Figure 13 has indicated that the user
has initiated the system shutdown process and provides a
performance summary for the last 100 processed frames.
The average collision probability is 0.098. It indicates a
low risk of collision, with an average distance to detected
objects of 315.62 cm. Each frame has taken an average
of 0.032 seconds to process, using 47.8% of available
RAM (Total RAM is 4GB). Since startup, 3600 frames
have been processed, resulting in 5 collision alerts. The
final entry confirms that the system has shut down safely.

FIGURE 14. Diagram for retrieving latency and response time.

TABLE 9. Latency and response time for AWS Greengrass integration.

FIGURE 13. CLI Output for the specific timestamp with the system closing.

E. AWS PERFORMANCE ANALYSIS

The performance of AWS Greengrass in our Safe-
Trax system has been assessed using several metrics.
It includes latency, response time, data throughput, and
resource utilization. A detailed analysis is given below.
The outputs from the CLI have been structured in a 2) DATA THROUGHPUT
tabular format. The commands have been executed on Data throughput is crucial for handling high-definition
the Raspberry Pi using ‘greengrass-cli’. video and frequent sensor data. The camera has streamed

VOLUME 13, 2025 6679

 M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

images to the Greengrass Core, where each frame bandwidth peaks at 500 KB/s during high-load tasks that
is processed in real-time by the Pre-Collision Alert show stable performance and effective management.
System. Performance metrics, like data throughput
(MB/s) and frames processed per second (FPS), have
been logged. The Greengrass Core has retrieved and
displayed logs from the last 15 minutes. The process for
how AWS Greengrass retrieves Data throughput using a
CLI command from the user-end has been presented in
figure 15. The output results are given in table 10.

FIGURE 16. Diagram for retrieving resource utilization.

TABLE 11. System performance metrics for AWS Greengrass deployment.

FIGURE 15. Diagram for retrieving data throughput.

TABLE 10. Performance metrics of the pre-collision alert system.

4) RELIABILITY AND FAULT TOLERANCE

The AWS Greengrass Core monitors the network, data
integrity, and system stability. In case of a failure,
An average throughput of 2.5 MB/s and peak of it recovers, stores local data, and logs metrics like
3.8 MB/s show the system handles complex scenes well. recovery success rate, recovery time, false alerts, and
Processing 25 FPS ensures real-time responsiveness for failover performance. Reliability metrics from the last
timely alerts and improved safety. 30 minutes have been retrieved and displayed in the CLI.
The process for how AWS Greengrass handles reliability
3) RESOURCE UTILIZATION and fault Tolerance using a CLI command from the
Resource utilization tracks AWS Greengrass overhead user-end has been presented in figure 17. The outputs
on the Raspberry Pi, including CPU, RAM, and band- are shown in table 12.
width. The Greengrass Resource Monitor has logged TABLE 12. AWS Greengrass system performance under network failure
average and peak values, which are processed and scenarios.
displayed in the CLI. The process for how AWS
Greengrass retrieves this information using a CLI
command from the user-end has been presented in
figure 16. The output results are given in table 11. The
system on AWS Greengrass efficiently uses Raspberry
Pi resources, with an average CPU at 37.5% (peaking at The table shows that AWS Greengrass recovers from
65%) and memory at 45.3% (peaking at 68%). Network network failures within 10 seconds, with no data loss

6680 VOLUME 13, 2025

M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

FIGURE 20. Training vs validation accuracy curves for MobileNetV2.

FIGURE 17. Diagram for retrieving reliability and fault tolerance.

or false alerts, indicating robust disconnection handling

and smooth failover management.

FIGURE 21. Training vs validation loss curves for MobileNetV2.

FIGURE 18. Training vs validation accuracy curves for ResNet50.

FIGURE 22. Training vs validation accuracy curves for VGG-16.

FIGURE 19. Training vs validation loss curves for ResNet50.

F. SAFETRAX INTEGRATED VEHICLE SAFETY SYSTEM:

PROTOTYPE DESIGN FIGURE 23. Training vs validation loss curves for VGG-16.
We have designed the SafeTrax integrated car prototype,
which is shown in Figure 9. It combines advanced sensing
and computing technologies to enhance vehicle safety. The detection and a dash cam mounted on the windshield or
system’s core components include a car with a sonar front of the vehicle for visual monitoring. These sensors
sensor installed in the bumpers and front grill for obstacle feed data to a Raspberry Pi, positioned discreetly within

VOLUME 13, 2025 6681

 M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

V. CONCLUSION AND FUTURE WORK

This section is divided into three parts consisting of our
overall discussion, conclusion, and our future work planning.

A. DISCUSSION
In this research, we have combined IoT and deep learning
technologies to create an effective accident prediction and
alert system. We have improved cloud services for in-car
sensors and cameras by using AWS IoT Greengrass. This
allows real-time data collection and processing. As a result,
the system’s speed, responsiveness, and ability to grow have
increased, making it easier to adapt to real driving conditions.
FIGURE 24. Confusion matrix for ResNet50.
To optimize performance, we have used and fine-tuned
several pre-trained deep-learning models in our SafeTrax.
This enables the system to handle complex situations,
including multi-dimensional data. We have also added an
alarm system to enhance safety. It provides drivers with
timely warnings about possible accidents. The system has
shown flexibility and it can work well in both urban and rural
areas, reducing the chances of accidents in different driving
conditions.

B. CONCLUSION
We have built a smart system that can predict car crashes and
collisions to help make roads safer. Our system uses some
available sensors which can be easily found everywhere, deep
FIGURE 25. Confusion matrix for MobileNetV2.
learning models, and AWS Greengrass to predict the probable
collision, access cloud services, and give drivers warnings in
real time if an accident might happen. The system is simple
and low-cost, and it works in different places. In our tests, the
system has worked very well. Two of our models, ResNet50
and MobileNetV2, have been very accurate and predicted
crashes correctly more than 93% of the time. Another model,
VGG-16, also did a good job, with over 90% accuracy.
We have also designed a prototype car using this SafeTrax
methodology. It has a car sonar sensor, a dashcam, a car
buzzer, and a Raspberry Pi. Images from each video frame
can be captured using car dashcams. The Raspberry Pi runs
crash-predicting models and connects to AWS Greengrass to
process these image data quickly. If there is a danger, a buzzer
FIGURE 26. Confusion matrix for VGG-16.
sounds to warn the driver. In a nutshell, SafeTrax can help
make roads safer and reduce car accidents. This smart and
the car, which serves as the central processing unit. The affordable system can be integrated into any type of vehicle.
Raspberry Pi runs deep learning models and AWS Greengrass In the future, we will keep improving the system to make sure
enabling both edge computing and cloud connectivity. This it meets all safety rules.
setup allows for real-time data processing and decision-
making. Python is used as the primary programming language C. FUTURE WORK
for the system. To alert the driver of potential hazards, To ensure road safety, we need to focus on enhancing
a car buzzer is installed near the dashboard or steering obstacle detection accuracy, enabling flexible real-time data
wheel. transmission for transportation systems, and ensuring scal-
This is the initial prototype of the SafeTrax car design, ability through the use of appropriate IoT sensors in future
which will be further developed in our future research. developments. To do this, we plan to add more convolutional
The prototype will require extensive testing and further layers and attention mechanisms. Since weather changes are
development to ensure its effectiveness and compliance with a major concern, we will create a weather-specific detection
automotive safety standards. module to help the system adapt to real-time weather

6682 VOLUME 13, 2025

M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

conditions and provide reliable predictions. Additionally, [14] N. Hu, D. Zhang, K. Xie, W. Liang, K. C. Li, and A. Y. Zomaya, ‘‘Dynamic
we will incorporate more advanced sensors. multi-scale spatial–temporal graph convolutional network for traffic flow
prediction,’’ Future Generat. Comput. Syst., vol. 158, pp. 323–332,
We will also integrate map-based applications. This will Sep. 2024.
give real-time updates on traffic patterns and promote safer [15] A. Razi, X. Chen, H. Li, H. Wang, B. Russo, Y. Chen, and H. Yu, ‘‘Deep
decision-making. Another key goal is to address sensor learning serves traffic safety analysis: A forward-looking review,’’ IET
Intell. Transp. Syst., vol. 17, no. 1, pp. 22–71, Jan. 2023.
quality constraints. We will focus on improving the quality
[16] C. Panda, A. K. Mishra, A. K. Dash, and H. Nawab, ‘‘Predicting and
of cameras and data processing. This will help ensure explaining severity of road accident using artificial intelligence techniques,
more accurate predictions, even in challenging situations, SHAP and feature analysis,’’ Int. J. Crashworthiness, vol. 28, no. 2,
pp. 186–201, Mar. 2023.
like low-quality image captures. Scalability challenges in
[17] N. Shatnawi, A. A. Al-Omari, and S. Alkhateeb, ‘‘Prediction of risk factors
remote areas with poor infrastructure can affect SafeTrax’s influencing severity level of traffic accidents using artificial intelligence,’’
performance. In the future, we will work on improving the Int. Rev. Civil Eng. (IRECE), vol. 14, no. 1, p. 1, Jan. 2023.
software to make it compatible with different networks. [18] Y. Ali, F. Hussain, and M. M. Haque, ‘‘Advances, challenges, and
future research needs in machine learning-based crash prediction models:
Lastly, to reduce processing delays in high-traffic situations, A systematic review,’’ Accident Anal. Prevention, vol. 194, Jan. 2024,
we will optimize AWS IoT Greengrass further. We will Art. no. 107378.
also increase the use of edge computing. This will help [19] X. Xiong, L. Chen, and J. Liang, ‘‘A new framework of vehicle collision
prediction by combining SVM and HMM,’’ IEEE Trans. Intell. Transp.
ensure faster and more responsive performance in various
Syst., vol. 19, no. 3, pp. 699–710, Mar. 2018.
situations. [20] A. V. Malawade, S.-Y. Yu, B. Hsu, D. Muthirayan, P. P. Khargonekar,
and M. A. A. Faruque, ‘‘Spatiotemporal scene-graph embedding for
REFERENCES autonomous vehicle collision prediction,’’ IEEE Internet Things J., vol. 9,
no. 12, pp. 9379–9388, Jun. 2022.
[1] R. Talbot, A. Filtness, and A. Morris, ‘‘Proposing a framework for [21] N. Venkateswaran, W. J. Hans, and N. Padmapriya, ‘‘Deep learning
evidence-based road safety policy-making: Connecting crash causation, based robust forward collision warning system with range prediction,’’
countermeasures and policy,’’ Accident Anal. Prevention, vol. 195, Multimedia Tools Appl., vol. 80, no. 14, pp. 20849–20867, Jun. 2021.
Feb. 2024, Art. no. 107409.
[22] L. Zhang, W. Xiao, Z. Zhang, and D. Meng, ‘‘Surrounding vehicles motion
[2] P. Vrtal, K. Kocián, J. Nováč ek, Z. Svatý, and T. Kohout, ‘‘A unified prediction for risk assessment and motion planning of autonomous vehicle
approach for risk assessment of road crash barriers using Bayesian in highway scenarios,’’ IEEE Access, vol. 8, pp. 209356–209376, 2020.
statistics,’’ Appl. Sci., vol. 13, no. 3, p. 1270, Jan. 2023.
[23] J. A. Onesimu, A. Kadam, K. M. Sagayam, and A. A. Elngar, ‘‘Internet
[3] R. Luke, ‘‘Current and future trends in driver behaviour and traffic safety
of Things based intelligent accident avoidance system for adverse
scholarship: An African research agenda,’’ Int. J. Environ. Res. Public
weather and road conditions,’’ J. Reliable Intell. Environ., vol. 7, no. 4,
Health, vol. 20, no. 5, p. 4290, Feb. 2023.
pp. 299–313, Dec. 2021.
[4] R. Ren, H. Li, T. Han, C. Tian, C. Zhang, J. Zhang, R. W. Proctor,
[24] L. Guo, Y. Jia, X. Hu, and F. Dong, ‘‘Forwarding collision assessment with
Y. Chen, and Y. Feng, ‘‘Vehicle crash simulations for safety: Introduction
the localization information using the machine learning method,’’ J. Adv.
of connected and automated vehicles on the roadways,’’ Accident Anal.
Transp., vol. 2022, May 2022, Art. no. 9530793.
Prevention, vol. 186, Jun. 2023, Art. no. 107021.
[25] C. Dong, C. Shao, J. Li, and Z. Xiong, ‘‘An improved deep learning
[5] V. Morozov, A. I. Petrov, V. Shepelev, and M. Balfaqih, ‘‘Ideology
model for traffic crash prediction,’’ J. Adv. Transp., vol. 2018, Dec. 2018,
of urban road transport chaos and accident risk management for
Art. no. 3869106.
sustainable transport systems,’’ Sustainability, vol. 16, no. 6, p. 2596,
Mar. 2024. [26] G. K. Sahoo, S. A. Patro, P. K. Pradhan, S. K. Das, and P. Singh, ‘‘An
[6] N. Sharma and R. D. Garg, ‘‘Real-time IoT-based connected vehicle IoT-based intimation and path tracing of a vehicle involved in road traffic
infrastructure for intelligent transportation safety,’’ IEEE Trans. Intell. crashes,’’ in Proc. IEEE-HYDCON, Apr. 2020, pp. 1–5.
Transp. Syst., vol. 24, no. 8, pp. 8339–8347, Aug. 2023. [27] L. Yu, B. Du, X. Hu, L. Sun, L. Han, and W. Lv, ‘‘Deep spatio-
[7] L. Muthugama, H. Xie, E. Tanin, and S. Karunasekera, ‘‘Real-time temporal graph convolutional network for traffic accident prediction,’’
road safety optimization through network-level data management,’’ Neurocomputing, vol. 423, pp. 135–147, Jan. 2021.
GeoInformatica, vol. 27, no. 3, pp. 491–523, Jul. 2023. [28] M. T. Kashifi, M. Al-Turki, and A. W. Sharify, ‘‘Deep hybrid learning
[8] Z. Ye, Y. Wei, S. Yang, P. Li, F. Yang, B. Yang, and L. Wang, ‘‘IoT- framework for spatiotemporal crash prediction using big traffic data,’’ Int.
enhanced smart road infrastructure systems for comprehensive real-time J. Transp. Sci. Technol., vol. 12, no. 3, pp. 793–808, 2023.
monitoring,’’ Internet Things Cyber-Phys. Syst., vol. 4, pp. 235–249, [29] D. Thapa, S. Mishra, N. R. Velaga, and G. R. Patil, ‘‘Advancing
Aug. 2024. proactive crash prediction: A discretized duration approach for predicting
[9] A. A. Musa, S. I. Malami, F. Alanazi, W. Ounaies, M. Alshammari, crashes and severity,’’ Accident Anal. Prevention, vol. 195, Feb. 2024,
and S. I. Haruna, ‘‘Sustainable traffic management for smart cities Art. no. 107407.
using Internet-of-Things-Oriented intelligent transportation systems (ITS): [30] M. M. Rathore, A. Paul, S. Rho, M. Khan, S. Vimal, and S. A. Shah,
Challenges and recommendations,’’ Sustainability, vol. 15, no. 13, p. 9859, ‘‘Smart traffic control: Identifying driving-violations using fog devices
Jun. 2023. with vehicular cameras in smart cities,’’ Sustain. Cities Soc., vol. 71,
[10] Y. Berhanu, E. Alemayehu, and D. Schröder, ‘‘Examining car accident Aug. 2021, Art. no. 102986.
prediction techniques and road traffic congestion: A comparative analysis [31] A. Pramanik, S. Sarkar, and J. Maiti, ‘‘A real-time video surveillance
of road safety and prevention of world challenges in low-income and high- system for traffic pre-events detection,’’ Accident Anal. Prevention,
income countries,’’ J. Adv. Transp., vol. 2023, Jul. 2023, Art. no. 6643412. vol. 154, May 2021, Art. no. 106019.
[11] F. A. Arnob, A. Fuad, A. T. Nizam, S. Barua, A. A. Choudhury, and [32] M. M. Islam, A. E. M. Ridwan, M. M. Mary, M. F. Siam, S. A. Mumu,
M. Islam, ‘‘An intelligent traffic system for detecting lane based rule and S. Rana, ‘‘Design and implementation of a smart bike accident
violation,’’ in Proc. Int. Conf. Adv. Emerg. Comput. Technol. (AECT), detection system,’’ in Proc. IEEE Region 10 Symp. (TENSYMP), Jun. 2020,
Feb. 2020, pp. 1–6. pp. 386–389.
[12] A. Kolhe, P. Potabatti, K. Khandagale, A. Jadhav, and J. Kulkarni, ‘‘Road [33] K. Dong, C. Zhou, Y. Ruan, and Y. Li, ‘‘MobileNetV2 model for image
companion-vehicle collision predication system,’’ vol. 6, no. 2, pp. 1–20, classification,’’ in Proc. 2nd Int. Conf. Inf. Technol. Comput. Appl. (ITCA),
Apr. 2018. Dec. 2020, pp. 476–480.
[13] T. Liu, A. Jiang, J. Zhou, M. Li, and H. K. Kwan, ‘‘GraphSAGE-based [34] S. Mascarenhas and M. Agarwal, ‘‘A comparison between VGG16,
dynamic spatial–temporal graph convolutional network for traffic predic- VGG19 and ResNet50 architecture frameworks for image classification,’’
tion,’’ IEEE Trans. Intell. Transp. Syst., vol. 24, no. 10, pp. 11210–11224, in Proc. Int. Conf. Disruptive Technol. Multi-Disciplinary Res. Appl.
Oct. 2023. (CENTCON), vol. 1, Nov. 2021, pp. 96–99.

VOLUME 13, 2025 6683

 M. Fahim Bin Amin et al.: SafeTrax: Smart Collision Prediction and Alert System Using IoT

[35] D. D. Hema and T. R. Jaison, ‘‘Efficient collision risk prediction model FARIA AKTER is a passionate researcher with
for autonomous vehicle using novel optimized LSTM based deep learning a BSc. in Computer Science and Engineering,
framework,’’ Int. J. Intell. Transp. Syst. Res., vol. 22, pp. 352–362, majoring in Data Science, from United Inter-
May 2024. national University. In her academic journey
[36] B. Alafi and S. Moradi, ‘‘Multi-class real-time crash risk forecast- she has gained hands-on research experience as
ing using convolutional neural network: Istanbul case study,’’ 2024, an undergraduate research assistant, where her
arXiv:2402.06707. work focused on advancing Machine Learning,
[37] S. Lee, B. Song, and J. Shin, ‘‘Collision prediction in an integrated
Natural Language Processing (NLP), and Human-
framework of scenario-based and data-driven approaches,’’ IEEE Access,
Computer Interaction (HCI). Through her projects,
vol. 12, pp. 55234–55247, 2024.
she has consistently applied her expertise to
address real-life rising challenges.
Currently, her research interests include Image Processing, Large
Language Models (LLMs), and Deep Learning. She aspires to broaden her
contribution further in the field of Data Science by solving real-life problems,
including different scenarios from several aspects.
MD. FAHIM BIN AMIN is a highly skilled
Machine Learning Engineer and Software Engi- ADIB AHMED received the degree in computer
neer with a robust background in building AI- science and engineering, specializing in software
powered systems and software solutions. With engineering from United International University
a degree in Computer Science and Engineering, (UIU), he focuses on creating practical and
majoring in Data Science from United Interna- impactful solutions that merge advanced technolo-
tional University, he has worked on innovative gies with real-world applications.
projects like vehicle number plate detection using He is dedicated to advancing research and
YOLOv8 and Alzheimer’s disease detection using innovation in the fields of machine learning, image
3D MRI images. processing, and the Internet of Things (IoT). With
His technical expertise spans a wide range of technologies, including a profound interest in artificial intelligence and
TensorFlow, PyTorch, React, [Link], and Docker, making him proficient in its transformative capabilities, he explores interdisciplinary approaches to
both software development and machine learning frameworks. His passion address complex problems. His work in AI and the IoT is complemented by
for innovation, combined with his expertise in both machine learning a strong emphasis on data science, as he employs data-driven methodologies
and software engineering, positions him as a forward-thinking engineer to devise innovative solutions. Driven by a passion for software development,
dedicated to driving technological progress. he aims to contribute to cutting-edge technological advancements and
promote significant progress within the technology sector.

MD. MOTAHARUL ISLAM (Member, IEEE)

received the Ph.D. degree in computer engineer-
ing from Kyung Hee University, South Korea,
ISRAT JAHAN KHAN is a Data Science major in 2013.
at United International University, with a strong He has been a Professor with the Department of
passion for Machine Learning (ML) and Deep Computer Science and Engineering, United Inter-
Learning (DL). Her academic research focuses national University (UIU), Dhaka, Bangladesh,
on computer vision and the development of since 2020. Before joining UIU, he worked at
hybrid ML algorithms, aiming to offer innovative many other universities at home and abroad,
solutions in these areas. such as the Islamic University of Madinah,
As a Machine Learning researcher, she has Saudi Arabia, the Islamic University of Technology, Brac University, and
gained valuable experience, further fueling her the University Grants Commission of Bangladesh. He has published around
passion for academia. Committed to advancing 100 articles more than the last ten years. His research interests include the
the field of data science, she strives to contribute to technological smart Internet of Things, IP-based wireless sensor networks (IP-WSN), WSN
progress through continuous learning, innovation, and impactful research in virtualization, cloud computing, and green computing.
education.

6684 VOLUME 13, 2025