Accurate Real-Time Traffic Light Detection Using YOLOv4

Abstract

Traffic light detection has a significant role in every driver assistance system (DAS). Using the information of recognized traffic signals, a DAS provides safe navigating suggestions to drivers. In this paper, we focus on horizontal traffic light formats, which are popular in South Korea, North American and most European countries. Despite their different shapes and sizes, we group important traffic light signals into six categories: red, yellow, green, green-turn, red-turn and no signal. This paper proposes a method to localize all traffic light positions in the current scene; and identify the main ones from them. To detect and classify traffic signals, we use YOLOv4, which is the most recent deep learning framework. Next, based on the relation of detected traffic lights, our system removes irrelevant signals. Finally, with the optimized set of detected traffic lights, we identify the main signal for the whole scene. Experiments were conducted on two large datasets captured in Seoul Korea containing both highways and urban areas. The results achieved 95% accuracy at 30FPS.