Real-time traffic sign recognition and autonomous vehicle control system using convolutional neural networks

Real-time traffic sign recognition and autonomous vehicle control system using convolutional neural networks

연구 분야: Artificial Intelligence

학회: Multimedia Tools and Applications

This paper presents the development of a real-time Traffic Sign Recognition (TSR) system integrated into an autonomous vehicle control platform. Leveraging deep Convolutional Neural Networks (CNNs), the system achieves high accuracy and robustness in classifying traffic signs, with an accuracy of over 99.63% for the German Traffic Sign Recognition Benchmark (GTSRB) and 99.68% for the Chinese Traffic Sign Dataset (CTSD). The image processing pipeline demonstrates consistent and efficient performance. When traffic signs are detected, the system achieves a latency of 24–26 ms per step, ensuring fast and reliable processing. Without traffic signs, the latency increases slightly, ranging from 24–30 ms per step, providing real-time performance suitable for dynamic driving environments. The CNN model is trained on the GTSRB dataset, which includes over 34,799 images across 43 categories of traffic signs, and CTSD comprises 6164 images across 57 classes. Key preprocessing techniques such as data augmentation (rotation, scaling, and translation), normalization, noise reduction through filtering, and grayscale conversion were applied to enhance the model's generalization capabilities. The TSR system operates in real-time, accurately detecting and classifying traffic signs captured by an onboard camera. Recognized signs directly inform the vehicle's control mechanisms, enabling dynamic adjustments such as speed regulation in response to speed limit signs and stopping upon detecting stop signs. The vehicle's control system is implemented using an Arduino microcontroller, integrating ultrasonic sensors for obstacle detection within 20 cm and infrared (IR) sensors for line following. A robust sensor fusion framework combines data from these sensors to enable dynamic decision-making. The IR sensors are primarily responsible for recognizing the road and initiating vehicle movement, ensuring the vehicle adheres to its designated path. Meanwhile, the ultrasonic sensors continuously monitor for obstacles in the environment. When an obstacle is detected, the ultrasonic sensor data takes precedence, causing the vehicle to stop and evaluate its surroundings, even if the IR sensors signal to continue moving. This integration of complementary sensor inputs ensures seamless navigation in traffic scenarios, resolving potential conflicts and prioritizing safety while maintaining adherence to road regulations. The proposed system offers a cost-effective and scalable solution for real-time autonomous driving applications. By combining advanced deep learning techniques with embedded systems, this work contributes a practical approach to implementing sophisticated driver-assistance features on low-cost platforms, advancing the field of autonomous vehicle technology.