Vehicle taillight detection and tracking using deep learning and thresholding for candidate generation

Vehicle taillights detection is an important topic in collision avoidance and in the field of autonomous vehicles. Analyzing the behavior of the front vehicle can prevent possible accidents. In this paper, a method for detecting vehicle taillights is presented. First, the system detects vehicles and then searches for candidate taillight pairs inside the obtained vehicles. Two methods for detecting candidate regions are presented. The first method uses explicit thresholds to extract red regions and the second method uses deep learning to segment taillights. Extracted candidates are then paired by comparing their sizes and centroid heights. Bhattacharyya coefficient is also used to validate taillight pairs by comparing their histograms. The system uses Kalman filtering to track detected taillights over time and to compensate for false negatives. The proposed solution is evaluated using the KITTI dataset.