Real-time detection and classification of traffic light signals

Traffic light detection is an important part of Advanced Driver Assist as well as autonomous vehicle systems which ensures timely and appropriate reaction to traffic lights (TLs) in cross sections. In this paper we introduce a robust and realtime approach to detect TLs and recognize its status in complex traffic scenes solely based on image processing techniques. The proposed system uses color properties of the scene to detect TLs in real-time. An innovative technique has been developed to significantly decrease compute requirement for detection of TL color by using one Lookup Table independent of lighting conditions. Each candidate region is further analyzed, using features analysis, to segregate actual TL signals among all candidate regions. As in similar machine learning techniques, an unsupervised classifier using a set of significant features has been developed to accurately segregate circular, semi-circular, and arrow shaped TL signals without using a training dataset. The final C++ code has been implemented and optimized on intelplatform using 1920×1080 frame resolution to recognize the status of TLs during day-time and night-time scenes, achieving 95% precision and 94.7% recall at 30FPS.