Improved YOLOv4-Based Object Detection Method for UAVs

Abstract

An improved UAV object detection method based on YOLOv4 is proposed in this paper for the problems faced by UAV vision detection, such as small targets, multiple scales, and complex backgrounds. First, in order to speed up the detection speed of the network and meet the actual detection demand, the backbone network is replaced with MobileNetv3 lightweight network, and the k-means++ is improved using a linear scale scaling method to improve the false detection rate by reclustering the prior frame; in addition, in order to reduce the loss of target information during downsampling, the stride convolution in PANet is replaced with non-stride convolution SPD-Conv, while further reducing the number of parameters and computational effort of the network model; for the small target of UAVs in the dataset, copy-pasting, a data enhancement strategy, is used to the UAVs to expand the dataset of small targets; finally, considering the problem that the complex background contributes significantly to the loss of the model, the Focal loss function is introduced, which interacts with the above methods to improve the accuracy and speed of the UAV detection model in complex backgrounds. The experimental results show that compared with the original YOLOv4, the proposed method improves the detection accuracy by 4.6%, the detection speed by 71%, and the missed detection rate by 17.9%, improving the UAV leakage problem in complex backgrounds while significantly improving the performance in terms of detection accuracy and detection speed.