Automatic and reliable identification of pedestrians from multiple camera views is very important for video surveillance and can save a lot of manual effort. The significant variations in viewpoints, poses, illumination and occlusions makes this problem very challenging. Most of the existing approaches addressing this problem handle drastic viewpoint change in a supervised way and thus require labelling new training data for a different pair of camera views. In this paper, we present a novel approach for pedestrian re-identification using stereo matching, which does not require any kind of training. The cost of the stereo matching of two images is used for evaluating the similarity of the images, without performing 3-D reconstruction. We show that this cost is robust to the large pose variations observed in the images captured from multiple cameras. The proposed pedestrian re-identification algorithm is built on top of a dynamic programming stereo matching algorithm. Experimental evaluation on the challenging VIPeR dataset shows the effectiveness of the proposed approach.
展开▼
