Non-overlapping Multi-Camera Object Tracking

With the development of computer visual, single camera intelligent visual surveillance algorithms have been mature step by step, object tracking in multi-camera network gradually become the key issue in recent years. Considering the contradiction between the broadness of the surveillance areas and the limited view range of the signal camera, the computation amount and the economical efficiency, it is impossible for the cameras to cover the whole region to be watched. So object tracking, especially non-overlapping multi-camera object tracking become an important part of research in broad area video surveillance.In non-overlapping multi-camera video surveillance system, object is separated both in temporal and spatial. How to match the object from different cameras is the core issue of the non-overlapping multi-camera object tracking. In order to settle this problem, we propose a multi-feature data confusion algorithm which combines several object appearance features with the topology relationship among the cameras. The main job and contributions in this paper are list as follows:(1) As the matching problem of object appearance model in non-overlapping multi-camera system, multiply appearance features and their matching algorithm are proposed including fragment histogram corrected by bright transfer function, UV chrome model and highness of the objects. As to the problem of gray distortion of the object among multi-camera, we use brightness transfer function to reduce the imaging difference caused by the camera itself and the circumstance illumination, then, cut the object model into fragment to make up for the space information loss of the regular histogram. The result shows that the fragment histograms corrected by the brightness transfer function rise the accuracy of the matching. For the reason that the UV chrome model is obtuse to the change of circumstance illumination, a kind of UV chrome model is used to match the object. It loses the matching error come from the illumination change. Due to the difference of the visual angle and the focal distance of the cameras, the same object images formed in different cameras has an obvious discrepancy on object highness. But in the certain circumstances, the object highnesses from two cameras satisfy a certain linear discipline in a nearly way. So we use the highness transfer model to exclude some outsides. The three matching model complete each other mutually, and improve the accuracy of the appearance matching enormously. (2) Aimed at the drawback that learning the topology relationship of the multi-cameras network is very time consumed, so a rapid online topology obtain method is proposed in this paper. The exit and entry are detected according to the tracking result in single camera, and single Gaussian model is employed to describe the probability distribution of the translation time for a single object, and the cumulative method is used to maintain the probability distribution of the translation time between cameras, which gives a responsible time clue for the object matching.(3) Considering the character of non-overlapping multi-camera object tracking system, the D-S evidence theory is applied to merge the multi-appearance models and the time-space restriction. This algorithm has tracked the object through the invisible region successfully; it also has avoided the conflicts among the evidences and creased the precision of the tracking.