An Improved Lightweight Network MobileNetv3 Based YOLOv3 for Pedestrian Detection
Xiaxia Zhang, Ning Li, Ruixin Zhang
Abstract
Recently, most object detection under videos have increasingly relied on the Unmanned Aerial Vehicle (UAV) platforms because of UAVs' timeliness, pertinence, and high flexibility in data acquisition. Convolution neural networks, especially for YOLO v3, have proved to be effective in intelligent pedestrian detection. However, two problems need to be solved in pedestrian detection of UAV images. One is more small pedestrian objects in UAV images; the other is the complex structure of Darknet53 in YOLO v3, which requires massive computation. To solve these problems, an improved lightweight network MobileNetv3 based on YOLO v3 is proposed. First, the improved MobileNetv3 takes place of the Darknet53 for feature extraction to reduce algorithm complexity and model simplify. Second, complete IoU loss by incorporating the overlap area, central point distance and aspect ratio in bounding box regression, is introduced into YOLO v3 to lead to faster convergence and better performance. Moreover, a new attention module SESAM is constructed by channel attention and spatial attention in MobileNetv3. It can effectively judge long-distance and small-volume objects. The experimental results have shown that the proposed model improves the performance of pedestrian detection of UAV images.