Object tracking is the task of approximating the trajectory of an object as it moves in a certain area. In most cases, covering the complete surveillance area is not feasible by a single camera due to limitations of cameras field of view (FOV) and the structure of scene. Therefore, surveilling a wide area requires a system with the ability to track objects across multiple cameras. Moreover, it is not feasible to completely cover wide areas, such as shopping centers, by an overlapping camera network due to the structure of the environment, and computational constraints. Thus, in realistic scenarios, the system should be able to handle multiple cameras with non-overlapping fields of view. Each camera observes a disjoint part of the environment that Tracking in such situation requires camera-to-camera correspondence of observations (I.e. re-identification of an object) when an object leaves one scene and appears at another scene later. Tracking people as they move through a camera network with non-overlapping field of view is a challenging issue.
In this project ( My M.Sc thesis) a multi-camera tracking system is proposed, where a number of stationary cameras are installed at an indoor environment and have non-overlapping fields of view. Tracking in such situation requires re-identification of objects when they leaves one field of view and appears at another scene later. Tracking in such systems is a challenging issue due to the changes of their appearance among the cameras. Appearance of a pedestrian is represented by its color histogram that is the consequence of illumination changes, parameters of camera, viewing angle and materials of the clothing. In addition in real world multi camera systems, the observations of people are often occurred arbitrarily in a widely separated time and position, and common prediction techniques such as Kalman filter cannot be used to establish correspondence between observations from different cameras.