End-to-End Person Re-identification including Camera Zooming based on Meta-Analysis for Images

This paper proposes an end-to-end person re-identification method including pan-tilt-zoom camera control in a real-world environment containing many moving people. Person re-identification, which means consistently identifying the same person from images from multiple cameras, brings great benefits to many services such as real-time person search. However, re-identification in the real world has challenges such as uniqueness of appearance, image quality, lighting, and poses of person. We address these challenges in two ways. First, we use the face images for re-identification because the face is the most unique feature of a person. Second, to acquire high-resolution images for facial recognition, our approach utilizes camera zooming. On the other hand, the reduction of field of view by zooming decreases the number of probed people. This paper proposes a method that resolves this trade-off by selecting the probed person from many people in the environment. Although many prior works have focused on data analysis by assuming data are pre-given, the proposed method solves the joint problem of facial recognition and image acquisition. The method selects the probed person who is likely to generate effective images for improving the accuracy of re-identification on the basis of the recognition results and the resolution, quantity, and variation of collected images. Experiments using a video dataset showed that the proposed method increased the success rate of re-identification by 25% compared with random selection.