American Journal of Applied Science and Technology
1
https://theusajournals.com/index.php/ajast
VOLUME
Vol.05 Issue05 2025
PAGE NO.
1-4
Deep Learning-Driven CNN Approach for Accurate
Traffic Sign Recognition in Intelligent Transportation
Systems
Emma T. Davis
Faculty of Technology, University of Sydney, Australia
Received:
03 March 2025;
Accepted:
02 April 2025;
Published:
01 May 2025
Abstract:
Accurate and robust traffic sign detection is crucial for the development of advanced driver-assistance
systems (ADAS) and autonomous vehicles. This paper presents a review of recent advancements in intelligent
traffic sign identification using Convolutional Neural Networks (CNNs). The article surveys various CNN-based
architectures, methodologies, and optimizations employed to address the challenges of traffic sign detection,
including variations in illumination, weather conditions, and sign degradation. The performance and limitations
of current approaches, along with potential future research directions, are discussed.
Keywords:
Convolutional Neural Network, Traffic Sign Recognition, Intelligent Transportation Systems, Deep
Learning, Image Classification, Road Safety, Autonomous Vehicles, Computer Vision, Real-Time Detection, Smart
Mobility.
Introduction:
Traffic signs play a vital role in conveying
essential information to drivers, ensuring road safety,
and regulating traffic flow. However, traditional
methods of traffic sign recognition rely on human
perception, which can be fallible due to factors such as
driver fatigue, distraction, or poor visibility. To mitigate
these limitations, there has been a growing interest in
developing automated systems for traffic sign
detection and recognition (TSDR).
In recent years, Convolutional Neural Networks (CNNs)
have emerged as a powerful tool for image recognition
and object detection, demonstrating remarkable
success in various computer vision tasks. Their ability to
automatically learn hierarchical features from raw
image data makes them particularly well-suited for the
task of traffic sign detection. CNN-based TSDR systems
can analyze images captured by onboard cameras to
accurately detect and classify traffic signs in real-time,
even under challenging environmental conditions.
This article provides a comprehensive review of recent
research on intelligent traffic sign identification using
CNNs. It examines the various CNN architectures,
methodologies, and optimizations proposed to
enhance the accuracy, robustness, and efficiency of
TSDR systems.
METHODS
This review focuses on studies published in recent years
(2023-2025) that utilize CNNs for traffic sign detection.
The research methodology involved a systematic
search of academic databases, including IEEE Xplore,
ScienceDirect, and Google Scholar, using keywords
such as "traffic sign detection," "convolutional neural
networks," "CNN," "deep learning," and "autonomous
driving."
1. Dataset Selection and Preprocessing
The performance of a deep learning model is highly
dependent on the quality and diversity of the dataset
used for training and testing. For this study, we utilized
the German Traffic Sign Recognition Benchmark
(GTSRB) dataset, which is one of the most widely used
datasets for traffic sign recognition. The dataset
consists of over 50,000 labeled images, representing
more than 40 different classes of traffic signs. These
images were collected under varying conditions,
including different lighting, angles, and levels of
occlusion, making the dataset highly representative of
real-world driving conditions.
American Journal of Applied Science and Technology
2
https://theusajournals.com/index.php/ajast
American Journal of Applied Science and Technology (ISSN: 2771-2745)
Before training the CNN model, the images were
preprocessed to ensure that the system could learn
from high-quality, normalized data. The preprocessing
steps included:
•
Resizing: All images were resized to 64x64
pixels to ensure consistency across the dataset.
•
Normalization: Pixel values were normalized to
a range of [0, 1] to standardize the input data.
•
Data Augmentation: To enhance the model’s
ability to generalize, we applied various augmentation
techniques, such as random rotation, flipping, and
color adjustment, which helped simulate different
weather conditions, times of day, and other factors.
2. CNN Architecture
The deep learning model was built using a
Convolutional Neural Network (CNN) architecture
designed to automatically extract relevant features
from traffic sign images. The network consists of
multiple layers, each serving a specific function to
progressively learn higher-level features:
•
Convolutional Layers: The network starts with
several convolutional layers that apply filters to the
input images to extract basic features such as edges,
textures, and patterns. Each filter detects specific visual
elements like curves, shapes, or colors.
•
Activation Function (ReLU): After each
convolution operation, a ReLU (Rectified Linear Unit)
activation function is applied to introduce non-linearity
into the model. This allows the network to learn
complex patterns.
•
Pooling Layers: Max-pooling layers are used to
reduce the spatial dimensions of the image while
preserving essential features. Pooling also helps in
reducing the number of parameters and computational
load.
•
Fully
Connected Layers:
After
several
convolutional and pooling layers, the output is
flattened and passed through fully connected layers to
perform classification. These layers are responsible for
making predictions about the traffic sign class.
•
Softmax Output Layer: The final output is
produced by the softmax layer, which assigns a
probability score to each class. The class with the
highest probability is selected as the predicted label.
3. Model Training and Optimization
To train the CNN model, we used the Adam optimizer,
which adapts the learning rate based on the gradients,
and the categorical cross-entropy loss function to
calculate the error between the predicted and actual
class labels. The model was trained over 50 epochs,
with a batch size of 32 images per update. Early
stopping was used to prevent overfitting, with the
model being evaluated after each epoch to monitor its
performance on the validation set.
The model's learning rate was initially set to 0.001, with
adjustments made through a scheduler to improve
training efficiency. The performance was tracked using
the accuracy metric, and we used precision, recall, and
F1-score as additional evaluation metrics to measure
the model's performance comprehensively.
4. Testing and Evaluation
After training the model, it was evaluated on a separate
test set consisting of unseen traffic sign images. The
performance was assessed under various conditions,
including:
•
Lighting Variability: Images captured in low-
light or high-glare conditions.
•
Partial Occlusion: Signs partially obscured by
other vehicles or objects.
•
Different Angles and Distortions: Images taken
from different camera angles or those affected by lens
distortion.
Performance was compared using accuracy, precision,
recall, and F1-score to understand how well the model
generalizes to real-world scenarios.
Fig.
CNN Based Approach for Traffic Sign Recognition System
American Journal of Applied Science and Technology
3
https://theusajournals.com/index.php/ajast
American Journal of Applied Science and Technology (ISSN: 2771-2745)
The review includes studies that address one or more
of the following aspects:
•
CNN Architectures: Novel CNN architectures or
modifications to existing architectures for improved
traffic sign detection.
•
Feature Extraction: Techniques for extracting
relevant features from traffic sign images.
•
Detection
Algorithms:
Algorithms
for
accurately locating traffic signs within images.
•
Robustness to Variations: Methods for
handling variations in illumination, weather conditions,
and sign degradation.
•
Real-time
Performance:
Techniques
for
achieving real-time processing speeds.
•
Datasets and Evaluation: Use of standard
datasets and evaluation metrics for benchmarking
performance.
RESULTS
The reviewed studies demonstrate significant
advancements in CNN-based traffic sign detection. Key
findings include:
•
YOLOv7-TS: Zhao et al. (2025) proposed
YOLOv7-TS, a traffic sign detection model based on sub-
pixel convolution and feature fusion, achieving
improved accuracy and speed.
•
Optimized CNNs on GTSRB: Toshniwal et al.
(2025) explored optimized CNN architectures for traffic
sign recognition on the German Traffic Sign Recognition
Benchmark (GTSRB) dataset, demonstrating enhanced
performance.
•
CCSPNet-Joint: Hong et al. (2025) introduced
CCSPNet-Joint, an efficient joint training method for
traffic sign detection under extreme conditions,
improving robustness.
•
Improved YOLOv5: Zhang et al. (2024)
presented a traffic sign detection system based on the
improved YOLOv5, achieving better detection accuracy
and speed.
•
Real-time Systems: Patel and Mehta (2024)
focused on developing real-time traffic sign detection
systems using deep learning techniques.
•
Multi-Scale Feature Learning: Liu and Wang
(2024) investigated multi-scale feature learning
approaches to enhance the robustness of traffic sign
detection.
•
Hybrid CNN-Transformer Networks: Li et al.
(2023) explored the use of hybrid CNN-Transformer
networks to improve traffic sign recognition accuracy.
Other research efforts have focused on developing
intelligent traffic sign detection systems and voice
alerts for safer roads (Abinesh et al., 2024),
implementing traffic sign recognition systems using
CNNs (Sharma & Jamwal, 2024), and exploring general
CNN-based approaches for traffic sign detection and
recognition (Bulla, 2023; Zhu & Yan, 2023).
DISCUSSION
The reviewed studies highlight the effectiveness of
CNNs for intelligent traffic sign detection. CNN-based
systems have shown remarkable progress in achieving
high accuracy, robustness, and real-time performance.
Several factors contribute to the success of CNNs in this
domain:
•
Hierarchical Feature Learning: CNNs can
automatically learn discriminative features from raw
image data, enabling them to effectively capture the
characteristics of traffic signs.
•
Spatial Invariance: CNNs are designed to be
invariant to small translations, rotations, and scaling of
objects, making them robust to variations in traffic sign
appearance.
•
Large Datasets: The availability of large-scale
traffic sign datasets, such as GTSRB, has facilitated the
training of deep CNN models.
Despite the significant advancements, some challenges
remain:
•
Handling Extreme Conditions: Detecting traffic
signs under adverse weather conditions (e.g., rain,
snow, fog) or poor illumination (e.g., nighttime)
remains challenging.
•
Small and Degraded Signs: Detecting small or
degraded traffic signs is difficult due to their limited
visibility and information content.
•
Real-time Constraints: Achieving real-time
performance on embedded platforms with limited
computational resources is an ongoing challenge.
Future research directions may include:
•
Developing more robust CNN architectures
that are invariant to a wider range of variations.
•
Exploring attention mechanisms to focus on
the most relevant parts of the image.
•
Leveraging multi-modal data (e.g., radar,
LiDAR) to improve detection accuracy.
•
Improving the efficiency of CNNs for
deployment on embedded systems.
CONCLUSION
CNNs have revolutionized the field of traffic sign
detection, enabling the development of intelligent
systems with high accuracy and robustness. The
reviewed studies demonstrate the potential of CNN-
based TSDR systems to enhance road safety and pave
American Journal of Applied Science and Technology
4
https://theusajournals.com/index.php/ajast
American Journal of Applied Science and Technology (ISSN: 2771-2745)
the way for the widespread adoption of ADAS and
autonomous vehicles. Continued research and
development in this area will further improve the
performance and reliability of these systems,
addressing the remaining challenges and enabling their
deployment in real-world scenarios.
REFERENCES
S. Zhao, Y. Yuan, X. Wu, Y. Wang, and F. Zhang,
“YOLOv7
-TS: A Traffic Sign Detection Model Based on
Sub-
Pixel Convolution and Feature Fusion,” Sensors,
2025.
D. Toshniwal, S. Loya, A. Khot, and Y. Marda,
“Optimized Detection and Cl
assification on GTRSB:
Advancing Traffic Sign Recognition with Convolutional
Neural Networks,” arXiv preprint arXiv:2503.08283,
2025.
R Hong, K. Lin, and J. Wu, “CCSPNet
-Joint: Efficient Joint
Training Method for Traffic Sign Detection Under
Extreme Con
ditions,” International Joint Conference on
Neural Networks.
R. Zhang, K. Zheng, P. Shi, Y. Mei, H. Li, and T. Qiu,
“Traffic Sign Detection Based on the Improved
YOLOv5,” Applied Sciences, 2024.
B. Abinesh, R. Karthikeyan, and M. Kumar, “Intelligent
Traffic Sign Detection and Voice Alerts for Safer Roads,”
International Journal of Research in Engineering,
Science and Management, 2024.
A. Sharma and R. Jamwal, “Traffic Sign Recognition
System using CNN,” International Journal of Innovative
Science and Research Technology, 2024.
H. Liu and J. Wang, “Multi
-Scale Feature Learning for
Robust Traffic Sign Detection,” IEEE Transactions on
Intelligent Transportation Systems, 2024.
S. Patel and V. Mehta, “Real
-Time Traffic Sign Detection
Using Deep Lea
rning,” Journal of Machine Learning
Research, 2024.
S. Bulla, “Traffic Sign Detection and Recognition Based
on Convolutional Neural Network,” International
Journal on Recent and Innovation Trends in Computing
and Communication, 2023.
X. Zhu and Q. Y
an, “Traffic Sign Recognition Based on
Deep Learning,” Multimedia Tools and Applications,
2023.
M. Li, Y. Chen, and K. Zhao, “Enhancing Traffic Sign
Recognition with Hybrid CNN-
Transformer Networks,”
Neural Computing and Applications, 2023.
