People Re-identification in Non-overlapping Field-of-views using Cumulative Brightness Transform Function and Body Segments in Different Color Spaces

Non-overlapping field-of-view (FOV) cameras are used in surveillance system to cover a wider area. Tracking in such systems is generally performed in two distinct steps. In the first step, people are identified and tracked in the FOV of a single camera. In the second step, re-identification of the people is carried out to track them in the whole area under surveillance. Various conventional features such as clothes and appearance of person have been used to identify peoples in the cameras. However, similarity between appearance and clothes of people causes unreliable results. Thus, much reliable features are still required to increase the ability of tracking system. The aim of this paper is to propose a novel algorithm to identify people in a network of cameras with disjoints views. In our proposed methodology, according to relative size of various parts, human body are divided into three portions of head, middle section and lower section. Considering histograms of these portions, matching process was done aiming to distinguish persons in FOV of cameras. The other main challenge in such multi-camera system is to deal with illumination changes and varying appearance of people at different cameras view. The cumulative brightness transform function (CBTF) was employed to alleviate the difficulty of illumination changes. Experiments were conducted in RGB, YCbCr and HSV color spaces and it was found that YCbCr color space gives a better performance result compared to other color spaces used in this work.