Traffic Light Recognition with Prior Maps

Traffic lights are crucial for urban cities. While it seems like an intuitive task for drivers to differentiate the status of the traffic light relevant to their path, it still poses a challenge for autonomous vehicles. Moreover, advanced driver assistance systems (ADAS) can also find this task challenging to warn inattentive drivers. Deep learning for traffic light detection and recognition is a popular solution to this problem. However, the problem arises when there are multiple traffic lights in the scene with different states Where choosing the correct traffic light to follow becomes a real challenge, which can be solved by using prior maps. In this paper, a two-fold solution is proposed. First, a mobile application is created that constructs prior maps for traffic lights autonomously. Second, the constructed prior maps and deep learning approaches are used in conjunction with computer vision approaches to assist the driver in determining the status of the relevant traffic lights. In this paper, YOLO- v4 was utilised. The proposed framework was trained on the LISA traffic light dataset, while the prior maps and testing were conducted on the streets of Cairo. It was shown that the proposed approach was able to correctly and promptly identify the status and relevancy of the traffic lights with 92.29% mean Average Precision (mAP).