关键词: YOLOv5s, 小目标, 车辆检测, K-means++, SPD-Conv, 检测头解耦模块

Abstract: The problem of complex backgrounds and large variations in target scales in vehicle images captured by unmanned aerial vehicle （UAV） makes it difficult for existing neural network models to detect small target objects when performing vehicle detection， which can easily lead to false detection and missed detection of small target objects. To solve this problem， an improveed method based on the YOLOv5s neural network is proposed. Firstly， we use the K-means++ algorithm to cluster dataset to obtain better anchor. Secondly， the SPD-Conv small target detection module is combined to reduce the false detection and miss detection rate， so as to improve the accuracy of vehicle detection. Finally， the detection head module is replaced by a decoupled head module to decouple the classification and regression tasks， thus further improve the classification accuracy. The article uses VisDrone-2019-DET dataset for vehicle detection， the mean average precision （mAP） of the improved network in this paper reaches 53.0%， which is 6.3 percentage points higher than the original YOLOv5s model， and can effectively reduce the probability of false detection and missed detection of small objects， enable more accurate vehicle detection.

Key words: YOLOv5s, small object, vehicle detection, K-means++, SPD-Conv, decoupled head model