Towards Efficient Traffic Light Detection and Attribute Recognition System in Open-World Complex Environments

Traffic lights are one of the most crucial elements of autonomous driving and smart cities to guarantee seamless and unobstructed road operation. With the motivation of designing redundancy mechanisms of cooperative vehicle infrastructure (CVI) system, this work aims to explore an effective and efficient traffic-light perception system based on vehicle-mounted cameras. Considering the scarcity of large-scale datasets for real-world scenarios in China, a traffic light detection and attribute (TL-DR) dataset is constructed. Additionally, a two-stage intelligent perception system, primarily consisting of lightweight detection and a multi-task network for color, arrow, and temporal attributes, is developed to guarantee stability and scalability. Considering the impact alterations in illumination and distance on detections, we optimized the generic detector YOLOv5-s to accommodate objects with small sizes at long distances. Furthermore, this work integrates the digital recognition of countdowns with the classification of color and orientation attributes into a multi-task framework to reduce the inference time. Extensive experiments conducted on the TL-DR database have demonstrated that the proposed system satisfies the real-time and high-precision requirements of vehicle safety while exhibiting good adaptability under challenging driving scenarios. The details will be available on https://github.com/pupu-chenyanyan/YOLO-TLR.