Person Re-identification using Kernel Learning and Similarity-distance Metric

Person re-identification involves matching pedestrian images observed from different cameras in non-overlapping multi-camera surveillance systems. In this article, a person re-identification method based on kernel-based similarity metric learning is proposed. First, the dimension of person fusion features is reduced by PCA whitening. Second, the kernel trick is used to deal with reduced features. Finally, the similarity metric function, which combines similarity and distance functions, is applied to the system; it helps the system learn affine transformation, which shows pairwise contrast. The result based on VIPeR, iLIDS, ETHZ, and CUHK01 datasets shows a significant improvement in performance measured in cumulative match characteristic curves. The proposed method is robust to different viewpoints, illumination changes, varying poses, and the effects of occlusion.