关键词: 交通灯检测, 注意力机制, 多尺度检测, 深度学习

Abstract: Aiming at the problems that the existing traffic light algorithm has poor detection and recognition effect on small targets and occlusive objects, this paper proposes a YOLOv5 detection algorithm based on feature fusion of attention and multi-scale (AM-YOLOv5). Firstly, by introducing a coordinate attention module into the residual block, the feature extraction ability of small targets is improved. Secondly, a four-scale detection layer is designed to improve the detection performance and accuracy of small-scale targets by introducing more shallower features. Finally, aiming at the problem that the introduction of attention and detection layers leads to an increase in the amount of computation and a decrease in speed, the method of replacing partial trunk convolution with distributed displacement convolution is used to simplify the model and improve the speed. The experimental results show that the average accuracy of the proposed algorithm reaches 90.8% in Lara dataset, which is 2.7% higher than that of the classical YOLOv5 algorithm, and the speed reaches 59.9 FPS. In the BDD100K dataset in a complex and harsh environment, the accuracy increases by 3.6%, and the speed reaches 34.8 FPS, which has a good detection effect and can better meet the real-time detection of traffic lights.