Deep learning for person re-identification in surveillance videos

Abstract

In the recent years, Closed Circuit Television (CCTV) is viewed as the basis for providing security. One of the most important aspects of CCTV surveillance systems security mechanism is to re-identify a person captured in one of the camera across different surveillance cameras. Re-identification has a major role in several applications like automated surveillance of universities, offices, malls, home and restricted environments like embassies or laboratories with strong security restrictions. Traditionally, identifying a person in a video was practiced under the set of same external conditions (like same illumination, viewpoint, back ground conditions etc.). But when it comes to automated re-identification in a CCTV surveillance system, several challenges emerge as the environment is uncontrolled and keeps varying, further the poses of the person and the angles of the cameras capturing the videos also incur additional challenge for the task considered. When a person disappears from one camera view for a period of time, he should be recognized in another view of camera at a different location when there are environmental disturbances like variation in illumination, crowded scene, partial occlusions, physical appearance variations, full occlusions, view point variations, background clutter, shadows and reflections, etc. In this chapter, the major focus is on the techniques of deep learning used to develop an end-to-end re-identification system highlighting the methods to handle the uncontrolled environment challenges mentioned. An end-to-end re-identification task consists of sequence of steps namely pedestrian detection, person tracking followed by person re-identification. Given a video sequence or an image as an input, firstly the humans are detected from the video sequence as a process of pedestrian detection. The person tracking within the camera is conducted, to find the different poses of the probe if needed. Then the re-identification process is conducted where the deep learning models are used to re-identify the person with the help of gallery set of videos and evaluates the similarities of gallery set and the person of interest by using deep learning metrics. The re-identification results end as a retrieval process where all similar images of the person of interest are retrieved. Several bench mark datasets considered in literature for re-identification system are VIPeR, ETHZ, PRID, CAVIAR, CUHK01, CUHK02, CUHK03, i-LIDS, RAiD, MARS, etc.