Object Tracking in Multiple Cameras with Disjoint

The specific problem we address in this chapter is object tracking over wide-areas such as an airport, the downtown of a large city or any large public area. Surveillance over these areas consists of the search for suspicious behavior such as persons loitering, unauthorized access, or persons attempting to enter a restricted zone. Until now, these surveillance tasks have been executed by human being who continually observe computer monitors to detect unauthorized activity over many cameras.It is known that the attention level drastically sinks after few hours. It is highly probable that suspicious activity would be unregistered by a human operator. A computer vision system, can take the place of the operator and monitor both immediate unauthorized behavior and long-term suspicious behavior. The "intelligent" system would, then, alert a responsible person for a deeper investigation. In most cases, it is not possible for a single camera to observe the complete area of interest because sensor resolution is finite and structures in the scene limit the visible areas. In the realistic application, surveillance systems which cover large areas are composed by multiple cameras with non overlapping Field of Views (FOVs). Multiple camera tracking is important to establish correspondence among detected objects across different cameras. Given a set of tracks in each camera we want to find which of these tracks belong to the same object in the real world. Note that tracking across multiple cameras is a challenging task because the observations are often widely separated in time and space, especially when viewed from non overlapped FOVs. Simple prediction schemes cannot be used to estimate the object position. It could be observed that object’s appearance in one camera viewmight be very different from its appearance in another camera view due to the difference in illumination, pose and camera properties. It is preferable that any multi-view tracking approach does not require camera calibration or complete site modelling. In fact, the benefit of calibrated cameras or site models is unavailable in most situations. Maintaining calibration between a large sensors network is a discouraging task, it is very onerous recalibrate any sensor when it (sensor) slight changes its position. This chapter presents an algorithm that deals with all described constraints and tracks objects across multiple un-calibrated cameras with non-overlapping FOVs. We investigate different techniques to evaluate intra-camera and inter-camera tracking algorithm based on object appearances. Object appearance can be modelled by its color or brightness, and it is a function of the scene illumination, object geometry, object surface material properties and camera Object Tracking in Multiple Cameras with Disjoint Views 11