Person Re-Identification by Weighted Integration of Sparse and Collaborative Representation

Recognizing the certain person of interest in cameras of different viewpoints is known as the task of person re-identification. It has been a challenging job considering the variation in human pose, the changing illumination conditions and the lack of paired samples. Previous matching techniques in the person re-identification field mainly focus on Mahalanobis-like metric learning functions. Taking advantage of the sparse representation and collaborative representation, we propose a new approach that elaborately exploits both the globality and locality of images. First, we explore multi-feature extraction with different spatial levels. The extracted features are then projected to a common subspace which handles dimension reduction. Second, we learn a single dictionary for each level that is invariant with the changing of viewpoints. Third, we adopt a weighted fusion approach that combines the dictionary learning-based sparse representation with collaborative representation. Experiments on two benchmark re-identification data sets (VIPeR and GRID) justify the advances of our integration algorithm by comparing with several state-of-the-art methods.