ABOS: An Attention-Based One-Stage Framework for Person Search

Abstract

Person search is of great significance to public safety research, such as crime surveillance, video surveillance and security. Person search is a method of locating and identifying the queried person from a complete set of images. The main cause of false recall and missed detection in person search is the presence of person occlusion in the images. In order to improve the accuracy of person search when the person to be queried is occluded, this paper proposes an attention-based one-stage framework for person search (ABOS) using an anchor-free model as a baseline. The method uses the channel attention module to express different forms of occlusion and take full advantage of the spatial attention module to highlight the target region of the occluded pedestrians. These attention modules integrate deep and shallow features to guide the network to pay attention to the visible area of the occluded target and extract the semantic information of the pedestrians. Experimental results on CUHK-SYSU and PRW datasets show that the proposed person search method based on attention mechanism in this paper has better performance than existing methods, achieving 93.7% of mAP on CUHK-SYSU dataset and 46.4% of mAP on PRW dataset, respectively