Density map guided vehicle detection on drone-captured scenarios

Vehicle detection on drone-captured scenarios is a challenging task because of the uneven distribution of vehicles, the huge variation in vehicle sizes and the real-time requirements. For detecting unevenly distributed and size-varying vehicles, inspired by the application of density map for crowd counting, a vehicle density-map estimation network (DENet) is proposed to learn the vehicle distribution information. Specifically, DENet is mainly composed of a low computation cost backbone (ResNet-18) and a lightweight Ghost-dilated Receptive Field Block (Gd-RFB). After density map generation, a simple region selection strategy is applied to form small crops for refined vehicle detection. Experimental results show that the integration of DENet to YOLOv5s can improve significantly the vehicle detection accuracy while guaranteeing real-time performance.