Vehicle Detection Using the Light Weight Model Based on YOLOv5

Abstract

Abstract Vehicle detection is important for the administration because it’s an important part of the intelligent traffic management system. It can help to make the best decision for the government and provide the data for intelligent roadside equipment. And can reduce road congestion and emissions, reduce the incidence of traffic accidents, improve road safety and reduce damage caused by accidents. The current method of object detection needs large computing resources and makes the roadside equipment costly or unable to meet the computing resources. So, this paper proposed a light weight detection model based on YOLOv5. This model reduces the parameter and FLOPs by using the deepthwise separable convolution CBAM to improve the performance of vehicle detection. And this paper reduces the input image size and halves the feature channel to reduce the amount of computation. The experiment result shows that the proposed model is 8 times smaller than YOLOv5-s while the mAP only decreases 6% on the UA-DETRAC dataset, significantly outperforming the other methods. This paper provided an idea for implementing a light weight model and a way to reduce the application cost of roadside smart devices.