CAE-Yolov5: A Vehicle Target Detection Algorithm Based on Attention Mechanism

Abstract

The number of vehicles has grown quickly along with people's quality of living and consumption, placing a significant stress on social transportation. The intelligent transportation system (ITS), which satisfies societal development requirements and lessens traffic congestion, was created. The primary area of research and its concentration have always been on accurate vehicle detection models. In this paper, we propose an improved Yolov5 network target identification algorithm for road vehicles utilizing attention mechanisms, aimed at the issues of complicated traffic scenes in the natural world, such as diverse vehicle sizes, overlapping vehicle targets, and substantial occlusion. First, a new attention module called CAE is developed that focuses the detection points of target detection on important areas by utilizing the benefits of convolutional block attention modules (CBAM) and efficient channel attention networks (ECA net). In order to collect shallower feature information, a larger-scale feature map is created, and the loss function of the bounding box, eiou (effective IOU loss), is increased. This accelerates convergence and enhances positional effects. According to experimental findings, the upgraded Yolov5 algorithm's average detection accuracy has increased from 91.3% to 96.3% when compared to the Yolov5 algorithm. On the Kitti dataset, the revised method outperforms the traditional one-stage SSD (single shot multi-box detector) and two-stage rapid RCNN (region CNN) algorithms in terms of detection accuracy and recall rate.