The goal of person re-identification (Re-id) is to match two pedestrian images of the same person which undergoes significant appearance changes in viewpoint, illumination and pose across camera views. To address this challenge, many algorithms have been proposed, and the researches can be divided into two major categories: feature extraction and metric learning. Most exciting studies of feature extraction focus on extracting robust features from pedestrian images directly. The extracted features include various hand-crafted color, shape, texture features, and some deeply learned features. As the second stage of person re-identification, metric learning plays an important role in learning a similarity function or a robust metric to optimize the matching score.

       However, various unfavorable factors such as changes in lighting, viewing angle, size, and partial occlusion remain challenges in designing a feature representation method which is discriminative, reliable and invariant to severe changes and misalignment across disjoint views. Also, the metric learning methods are affected by some classical problems, such as the inconsistent distributions between multiple views and the small-sample issue for model learning.

       In recent years, our laboratory has worked to resolve these issues. We have thoroughly studied the theory of feature extraction and metric learning, and proposed a series of solutions in  granular modeling of pedestrian image information and similarity measurement. The research results have been published in international academic journals and international academic conferences. At the same time, these research results have also been successfully applied in the smart campus system, and we have hosted two projects of the National Natural Science Foundation.