University of Houston

Person re-identification (Re-ID) is a fundamental task in automated video surveillance and has been an area of intense research in the past few years. Given an image/video of a person taken from one camera, Re-ID is the process of identifying the person from images/videos taken from a diffrent camera. Re-ID is indispensable in establishing consistent labeling across multiple cameras or even within the same camera to re-establish disconnected or lost tracks. Apart from surveillance it has applications in robotics, multimedia and forensics. Person Re-ID is a difficult problem because of the visual ambiguity and spatio-temporal uncertainty in a person's appearance across different cameras. In this work, we explore the problem of person Re-ID for multi-camera tracking, to understand the nature of Re-ID, constraints and conditions under which it is to be addressed and possible solutions to each aspect. We argue that Re-ID for multi-camera tracking is inherently a open set Re-ID problem with dynamically evolving gallery and open probe set. We propose multi-feature person models for both single and multi-shot Re-ID with a focus on incorporating unique features suitable for short as well as long period Re-ID. Finally, we adapt an open set recognition technique to address the problem of novelty detection within open set Re-ID. In conclusion we identify the open issues and challenges of Re-ID with a discussion on potential directions for further research.