Person re-identification using pose-driven body parts

The topic of Person Re-Identification (Re-ID) is currently attracting much interest from researchers due to the various possible applications such as behavior recognition, person tracking and safety purposes at public places. General approach is to extract discriminative color and texture features from images and calculate their distances as a measure of similarity. Most of the work consider whole body to extract descriptors. However, human body maybe occluded or seen from different views that prevent correct matching between persons. We propose in this paper to use a reliable pose estimation algorithm to extract meaningful body parts. Then, we extract descriptors from each part separately using LOcal Maximal Occurrence (LOMO) algorithm and Cross-view Quadratic Discriminant Analysis (XQDA) metric learning algorithm to compute the similarity. A comparison between state-of-the-art Re-ID methods in most commonly used benchmark Re-ID datasets will be also presented in this work.