Rider Re-identification Based on Pyramid Attention

Abstract

In recent years, object re-identification (ReID) based on deep learning has made great progress, and research in this field mainly focuses on person and vehicle. However, the researchers ignore an important target: riders. Electric bikes are an essential part of modern transportation scenarios, so identifying riders and monitoring their behavior on a large scale is critical to public safety management. To bridge the research gap of rider ReID, this paper proposes a pyramid attention network (PANet), which utilizes the pyramid structure to capture multi-scale clues and discovers key regional features from fine to coarse. PANet first learns fine-grained attention in local small regions, then gradually aggregates local regions to expand the scope of attention exploration, and finally conducts global coarse-grained attention learning. We implement two different dimensional attention computations in the pyramid attention network: spatial attention and channel attention. Experiments on BPReID and MoRe datasets demonstrate the effectiveness of this network design, which can achieve better performance with limited computational overhead.