A YOLOv5-based method for detecting vehicles in overhead surveillance video

As one of the core technologies of intelligent transportation system, vehicle detection technology has high requirements for its detection accuracy. For the characteristics of small vehicle targets and complex background in overhead surveillance video, this paper proposes a novel method by improving the original YOLOv5 algorithm. First, the very small target detection layer MSD (Minimal Scale Detection) is added to YOLOv5 network to improve the detection effect of YOLOv5 network on small target vehicles in overhead surveillance video. Second, the CA (Coordinate Attention) attention mechanism is added to the YOLOv5 network to further improve the detection accuracy. Third, the CIOU (Complete Intersection over Union) focal loss function used in the original YOLOv5 is replaced with EIOU (Efficient Intersection over Union) to accelerate the convergence of the prediction box and improve the regression accuracy of the prediction box. By testing on the self-built dataset, compared with the original YOLOv5 network, the mAP (mean Average Precision) value of the proposed method is improved by 3.01%, the Precision is improved by 1.33%, and the Recall is improved by 1.12%. The overall network performance is improved comprehensively, which is more suitable for the detection of small target vehicles in overhead surveillance video.