Person Re-Identification Based on Kernel Large Margin Nearest Neighbor Classification

Person re-identification is a process of matching person images of same identity across nonoverlapping camera views at different locations and times. In this paper, we introduce how to use the kernel trick to improve the performance of large margin nearest neighbor (LMNN) classification for person re-identification. Since the classification ability of LMNN is weak for those person features with nonlinear distribution, KLMNN combining kernel trick and LMNN is introduced to extend linear distance metric to nonlinear cases. Three kernel-based methods and two indicators are applied to evaluate the performance of KLMNN.