Pedestrian Detection Under Unmanned Aerial Vehicle an Improved Single-Stage Detector Based on RetinaNet

Chao Xiang, Haochen Shi, Ning Li, Meng Ding, Huiyu Zhou

Abstract: As a kind of single-stage detection network, RetinaNet has excellent performance in object detection, but it still has some difficulties in small object detection in unmanned aerial vehicle (UAV) images. Based on RetinaNet, this paper uses improved ResNet50 as backbone to extract features. We increase the number of anchor ratios, but reduce the number of anchor scales so that the number of anchors won't increase too much. We add the shallower features to predict smaller targets. We also add Receptive Field Block after each predicted feature to increase the receptive field. We have changed the values of parameters in the loss function for better detection performance. The improved model effectively increases the detection rate of small targets so that it can be applied to pedestrian detection under UAV. In our UAV pedestrian dataset, the improved model increases average precision (AP) from 51.29% to 84.05%.