Rahul Rama Varior, Asst. Prof. Wang Gang
Person re-identification deals with matching images of same person over multiple non-overlapping camera views. It is applicablein tracking a particular person across these cameras, tracking the trajectory of a person, surveillance, and for forensic and security applications.
- Multi-camera tracking: To track an individual using multiple cameras, the identity of the person has to be retrieved from the second camera based on the information obtained from the first camera.
- Tracking the trajectory: If the locations of the cameras are known, based on the re-identification system, it is possible to track the moving path of a person from one point to another. Eg: From Indoor office spaces, to shopping malls, to MRT stations, etc.
- Surveillance and security applications: To track the suspect of a crime scene, using multiple cameras. Intended for forensic applications.
Some example images from VIPeR dataset which is publicly available for research purposes:
Some of the major challenges for person re-identification system that is being addressed in the research community are
- Illumination changes: Day light intensity, shade, changes in illumination color in indoor and outdoor environments etc. This causes changes in the appeared color of the same subject across multiple camera views.
- Resolution: As many of the older surveillance cameras give low-resolution images, it is a difficult task for the algorithm to differentiate between individuals due to lack of “information”.
- Occlusion: In crowded scenes, it is difficult to obtain full-body image of an individual due to partial or full occlusion of that individual by others in the crowd.
- Uniform Clothing: For schools and some of the factory/construction sites, clothing is uniform for the subjects. If the algorithm is mainly based on the appearance, it is difficult to distinguish between subjects.
- Scalability: Public areas are equipped with thousands of cameras. Current technologies are only starting to address multi-camera surveillance problems.
- Lack of annotated data: Obtaining data is easy. But annotated data is always scarce. For good generalization capabilities, complex algorithms need to be taught with large number of labelled data.
Most of the re-identification research works are carried out focusing on two aspects of the problem. To develop a mathematical representation for the images, known as a feature representation, and to develop a distance function that can reduce the “distance” between individuals of same identity and increase the “distance” between individuals of different identity in an n-dimensional vector space.
- Illumination: An illumination invariant descriptor based on new technologies such as feature learning is proposed to handle appearance changes due to varying illumination.
- Labelled data: To tackle the problem of lack of annotated data, distance learning using marginalization is proposed.
A short demo of our results is given below.
 D. Gray, S. Brennan, and H. Tao. Evaluating appearance models for recognition, reacquisition, and tracking. In PETS, 2007. URL: https://vision.soe.ucsc.edu/node/178