Subspace learning for person Re-identification

Abstract

Person Re-Identi cation (PRID) has been one of the most prominent problems in intelligent video surveillance in the recent past. PRID is the task of retrieving all the instances of a person given a single shot or a sequence of images. Since the images come from non-overlapping camera views, the existing metric learning algorithms either try to learn a subspace or a metric or shared and private features. However, learning shared and private features are not explored in the subspace itself. In this thesis, we aim to develop a novel cross-view subspace learning method for extracting latent semantics from the hand-crafted features, and to thus perform PRID. We analytically derive the subspace as well as shared and private features. We explore the di erent variants of the proposed scheme and empirically analyze their performance.