Feature correspondence in a non-overlapping camera network

Abstract

Person re-identification across multiple cameras is difficult due to viewpoint and illumination variations. Most traditional research focuses on developing invariant features that are unaffected by these variations. However, thus far, there has been no feature developed that is completely invariant, and it is possible that a fully invariant feature may not exist. Therefore, we do not seek to develop these ideal features in this paper. We instead propose a framework for learning a gallery of persons who appear in the camera network frequently. The gallery contains appearance models of these persons from each camera and viewpoint. Given the camera identity, viewpoint identity, person identity, the model is decided. Since these appearance models are specific to each camera and viewpoint, the problems of viewpoint variations and illumination variations between cameras are explicitly solved, and re-identification becomes a ranking problem. Experiments demonstrate that our framework provides significant improvement in addressing the re-identification problem.