Vehicle detection on roads based on Yolov5 with multi-scale feature fusion

Abstract

This study focuses on Yolov5-based multi-head multi-scale adaptive feature fusion for vehicle detection to enhance the intelligence and refinement of road traffic safety management. As urbanization accelerates, road traffic problems are becoming increasingly serious. Accurate vehicle detection is crucial for traffic management to detect violations, monitor traffic flow, and prevent accidents in a timely manner. This paper proposes an improved Yolov5s-G model, which enhances the detection performance for small objects and improves the retention of feature information by introducing a small-object detection layer and a Weighted Cross-scale Fusion module (WCF), and an Adaptively Spatial Feature Fusion4(ASFF4) module. These enhancements enable the model to improve detection accuracy while maintaining moderate computational complexity. Specifically, the new small-object detection layer captures positional information of small objects more effectively, while the WCF module prevents the loss of small-object information during convolution through bidirectional cross-scale link feature fusion. Additionally, the ASFF4 module utilizes adaptive spatial feature fusion to further enhance the processing capability of feature information. Experimental results demonstrate that the improved Yolov5s-G model performs well on the vehicle detection dataset, with a mAP improvement of 9.3% compared to the original Yolov5 model. Furthermore, by introducing the knowledge distillation technique, the model has been significantly enhanced in terms of lightweighting.