Person Re-identification Based on Part Feature Specificity

Person re-identification has become one of the most important problems in video surveillance system. In a multi-camera video surveillance system with non-overlapped area, the appearances of one person are much difference according different cameras or in the same camera at different times. On the other hand, different person may appears similar in one camera, so which made person re-identification a challenging problem. This paper carried out a person re-identification algorithm based on part feature specificity. This algorithm extract color, texture and shape features of different part of body first, then gather statistic specificity weight of these features for each part. At last, doing feature weighting both part weight and feature specificity in distance calculating, which make features with strong specificity more important. This algorithm indicates some parts of body are more important than others in re-identification, and the same part from different people with different appearance, the features with strong specificity are more important than the others. We have done our experiments at public datasets VIPeR and iLIDS, and evaluate the result by CMC. Result indicates this algorithm has higher re-identification rate, and more robust to viewing condition changes, illumination variations, background clutter and occlusion.