Vehicle detection in surveillance videos based on YOLOv5 lightweight network

Abstract

The development of surveillance video vehicle detection technology in modern intelligent transportation systems is closely related to the operation and safety of highways and urban road systems. Yet, the current object detection network structure is complex, requiring a large number of parameters and calculations, so this paper proposes a lightweight network based on YOLOv5. It can be easily deployed on video surveillance equipment even with limited performance, while ensuring real-time and accurate vehicle detection. Modified MobileNetV2 is used as the backbone feature extraction network of YOLOv5, and DSC “depthwise separable convolution” is used to replace the standard convolution in the bottleneck layer structure. The lightweight YOLOv5 is evaluated in the UA-DETRAC and BDD100k datasets. Experimental results show that this method reduces the number of parameters by 95% as compared with the original YOLOv5s and achieves a good tradeoff between precision and speed.