A New Deep Learning Architecture for Person Detection

Abstract

Person detection is a branch of object detection. It refers to positioning people in the image, finding the location and range of the person, and has a wide range of applications in fields such as video surveillance and target tracking. Yolo3 is currently one of the best deep learning structure for object detection. In this paper we further improve the Yolo3 network by combining the excellent characteristics of the end-to-end network for person detection. We propose a new person detection network model called PDnet. Among the main contributions, we further optimize the Yolo3 feature extraction network structure, change the three output ports of Yolo3 to one, and improve the anchor boxe clustering algorithm, so that our network model can extract the person features better, speed up the convergence of the category loss in the original loss function. The experimental results show that compared to vannila Yolo3, our proposed PDnet has better robustness and higher accuracy in person detection.