Hybrid strategy for traffic light detection by combining classical and self‐learning detectors

Abstract

Detection of the traffic light is a key function of the automatic driving system for urban traffic. Considering the characteristics of classical and self-learning algorithms, a fusion logic is proposed to make up the shortcoming of learning algorithms by combining the known knowledge with the learning features to detect the red and yellow–green traffic light without turn indicator. The relationship of detection performance among different detectors is established analytically. Then the improvement of detection performance by fusion is analysed theoretically and optimised numerically. According to the analysis results, the hybrid detector is designed by using the colour information in hue-saturation-intensity to extract the candidate region, the hog feature to identify the shape information of traffic light classified by a support vector machine, and a comparatively simple convolutional neural network (CNN) with the classical AlexNet structure to act as the self-learned detector. The effectiveness of the hybrid method is validated by several comparative tests with single CNN detectors and other fusion methods on the training dataset, and the extensibility to new application conditions is evaluated by vehicle tests.