Traffic light detection for nighttime driving with log-polar transform incorporated learning

• An image-based method for traffic light detection and recognition is developed. • A two-stage processing framework with individual lights are detected for traffic state recognition. • Built upon on the light-priority network architecture, additional green arrow signals are included for model training. • The detection and classification networks are integrated into an end-to-end model with feature maps sharing. Recent advancements in advanced driver assistance systems (ADAS) have equipped vehicles with enhanced intelligence. Traffic light detection, a fun- damental function, is crucial for understanding transportation infrastructure. This paper addresses the challenging task of nighttime traffic light detec- tion, including arrow signal recognition. A learning-based approach with a two-stage framework is proposed to tackle nighttime driving scenarios. In the first stage, image intensity encoding and tone-mapping are employed for traffic light detection. Subsequently, an SVM classifier integrated with log- polar transformation is used to recognize arrow signal lights. Due to the scarcity of publicly available traffic light datasets specifically designed for nighttime scenes, a dedicated nighttime traffic light dataset is created for network training and testing. The proposed techniques are evaluated using real-world driving images. With the average accuracy of 97 . 5 % across all traf- fic light categories, the performance evaluation underscores the effectiveness and feasibility of our algorithms for nighttime driving scenarios.