An Anchor-free Detector Based on Residual Feature Enhancement Pyramid Network for UAV Vehicle Detection

Vehicle detection in Unmanned Aerial Vehicle (UAV) images is a challenging task because there are many small objects in UAV images, and the scale of objects varies greatly, which brings great difficulty to vehicle detection using existing algorithms. This paper proposes an anchor-free detector called Residual Feature Enhancement Pyramid Network (RFEPNet) for UAV vehicle detection. RFEPNet contains a Cross-Level Context Fusion Network (CLCFNet) and a Residual Feature Enhancement Module (RFEM) based on pyramid convolution. Specifically, CLCFNet utilizes the densely connected structure and Dual Attention Fusion Module (DAFM) to increase the sensitivity of high-resolution feature maps to small objects. Simultaneously, RFEM exploits pyramid convolution and residual connection structure to enhance the semantic information of the feature pyramid. In addition, the anchor-free head is used for classification and bounding box regression. The experimental results on the UAVDT dataset show that the proposed RFEPNet achieves state-of-the-art performance.