Development of Traffic Light and Road Sign Detection and Recognition Using Deep Learning

Traffic light and road sign violations significantly contribute to traffic accidents, particularly at intersections in high-density urban areas. To address these challenges, this research focuses on enhancing the accuracy, robustness, and reliability of Autonomous Vehicle (AV) perception systems using advanced deep learning techniques. The novelty of this study lies in the comprehensive development and evaluation of real-time traffic light and road sign detection systems, comparing state-ofthe-art models including YOLOv3, YOLOv5, and YOLOv7. The models were rigorously tested in a controlled offline environment using the Nvidia Titan RTX, followed by extensive field testing on an AV test vehicle equipped with sensor suite and Nvidia RTX GPU. The testing was conducted across complex urban driving scenarios at the CETRAN proving test track, JTC Cleantech Park, and NTU Singapore campus. The traffic light detection and recognition (TLR) results demonstrate that YOLOv7 outperforms YOLOv5 and YOLOv3, achieving a mean Average Precision (mAP@0.5) of 93%, even under challenging conditions like poor lighting and occlusions. While the traffic road sign detection (TSD) mAP@0.5 of 96%. This superior performance highlights the potential of YOLOv7 in enhancing AV safety and reliability. The conclusions underscore the effectiveness of YOLOv7 for real-time detection in AV perception systems, offering crucial insights for future research. Potential implications include the development of more robust and accurate AV systems, capable of safely navigating complex urban environments.