Person Re-identification Based on Part Feature Importance

In a video surveillance system, same person may looks different across different camera, while different people may look the same in one camera, thus making person re-identification a challenging problem. This paper carried out an algorithm based on part feature importance, firstly extract features such as color, texture and shape. Then cluster each part by classify different appearance of body parts, using an error accumulation method to figure out weight vector indicate the significant of the feature, which fit the type of appearance. Calculating the similarity at last, using this vector to weighting the features of each part, making the feature more suitable for the appearance. This algorithm indicate some features are more important than others for parts with different appearance. We had complete our experiments on the public datasets VIPeR, and evaluate the result by CMC curve, which indicated this algorithm achieved higher re-identification rate for re-identification, and more robust to viewing condition changes, illumination variations, background clutter and occlusion.