High-Altitude Pedestrian Detection Based on Improved YOLOv3

Qing Tian, Pengfei Cao, Haoyi Fan, Rong Hu, Zuoyong Li

Abstract: As one of the main tasks in the field of computer vision, pedestrian detection aims to find out all pedestrians in the image or video. The existing YOLOv3 is a relatively mature object detection method. However, for the long-distance pedestrian detection task in high-altitude scenes, YOLOv3 has the limitations of low detection speed and low detection accuracy. This paper proposes an improved YOLOv3 method briefly called YOLOv3-M for the high-altitude pedestrian detection, which replaces the feature extraction module called darknet53 in YOLOv3 with MobileNetv1. Specifically, YOLOv3-M first constructs the dataset with the small objects of high-altitude pedestrians as the detection object. Then, it uses the K-means +  + algorithm to re-cluster the high-altitude pedestrian dataset. Next, it uses the Distance Intersection over Union (DIoU) loss function to alleviate the problem of high-altitude pedestrian overlapping. Experimental results show that the proposed YOLOv3-M improves the detection precision and the detection speed compared to YOLOv3.