Research on person re-identification based on posture guidance and feature alignment

This paper proposes a person re-identification algorithm utilizing posture guidance and feature alignment to solve posture difference and misalignment of the retrieved pedestrian images. Our technique employs Openpose to locate 18 key points on the human body and integrates 18 heat maps of various human body key points into a global feature representation. Then, a hard attention mechanism based on the human body key points forces the network to focus on the pedestrian posture features to align the same body parts of pedestrian imagery. Our architecture solves the pedestrian image posture difference and misalignment problem and performs robust person re-identification. We challenge the developed method on the public Market1501 and DukeMTMC-reID datasets, employing the Rank-1 and mAP performance metrics, and obtain 94.6%/81.4% and 85.7%/72.7%, respectively. The results highlight that the proposed algorithm solves the problems of pedestrian image misalignment and posture difference, proving the effectiveness and practicability of the proposed algorithm.