Manifold-based learning for person re-identification

Abstract

The research described in this paper consists in developing a system for re-identifying people across multiple non-overlapping cameras. The proposed approach consists of three main steps: appearance-based feature extraction, data projection on manifold space, and similarity estimation for person re-identification. We first decompose the human image into a grid of patches and characterize each patch by a colorimetric feature vector. These patches are then embedded into a non-linear manifold, which preserves the local and global proximity among data points. Finally, a matching framework is introduced to estimate the similarity of image pairs and to make the final decision of re-identification. The performance of our system is evaluated on the well-known VIPeR dataset. The experimental results show that the proposed system leads to satisfactory results.