A YOLOv4-based Vehicle Detection Method from UAV Videos

Abstract

In order to solve the positioning problem and missed detection caused by occlusion or small target when the UAV detects the target vehicle, this paper proposes an end-to-end vehicle detection method based on UAV video. The shallow fine-grained information and high-level semantic information are extracted through the YOLOv4 detection network and feature fusion is performed, the loss function is partially replaced with Mish to optimize the flow of gradient information. The initial anchor box obtained by K-means clustering is used to realize a multi-scale detection strategy. Through testing in different environments, the proposed method has high robustness in terms of different vehicle scale and texture changes caused by UAV camera angle movement or weather changes. The mean Average Precision (mAP) index of the algorithm reaches 90.3%, which has good detection effect and positioning accuracy.