A Real-Time Vehicle Traffic Light Detection Algorithm Based on Modified YOLOv3

Abstract

Aiming at the problems of low recognition rate and high missed detection rate in deep learning algorithms for detecting traffic lights, as well as the scarcity of traffic light datasets in China. A real-time traffic light detection and recognition method based on the improved YOLOv3 algorithm is proposed. Firstly, the linear scale scaling method is used to optimize the aspect ratio of the prior box generated by K-means clustering, and the clustering result is linearly calculated to obtain a suitable anchor box size. Then, the improved Mosaic approach is used to enhance the traffic light dataset. Finally, in order to reduce the repeated feature extraction of the image by the convolutional neural network, a SPP block is added after the backbone network, and a 4 stride up-sampling layer is added to better integrate high-level semantic information and shallow location information. At the same time, the number of convolutional layers in the neck part is reduced, and the model structure is simplified. Experimental results show that the proposed approach achieves higher accuracy on both the Lara dataset and the Chongqing traffic light dataset (CQTLD), compared with YOLOv3 approach. The detection speed is increased by 11.8%, and mean Average Precision (mAP) is increased by 3.78% on CQTLD.