Research Of Key Problems Of Non-overlapping Multi-camera Object Tracking

In recent years, with the gradually mature of the infrastructural single camera intelligent visual surveillance algorithms, the research of the information associating and merging of non-overlapping multi-camera is given more attention. Considering the contradiction between the broadness of the surveillance areas and the limited viewing range of single camera, the computation amount and the economical efficiency, it is impossible to cover all of the areas to be surveillanced. As a consequence, Object tracking in the case of non-overlapping multi-camera is becoming an important part of the research of broad area video surveillance.In non-overlapping multi-camera video surveillance system, object is detached both in temporal and spatial. How to matching the object of different cameras is the critical problem of multi-camera object tracking. In this thesis, we proposed a method which uses the spatial-temporal relation together with the multi-performance model to realize object matching. The main work and contribution of this thesis is as follows:Firstly, as a basis for the research of object tracking in non-overlapping multi-camera situation, we do a deep research of single camera object tracking; realize the single Gauss background modeling, the graph-cut based object extracting and non-parametric kernel density estimation based color feature matching.Secondly, against the object-performance model matching problem between multiple cameras with no overlapping views, we propose a method which uses the color histogram together with the object height to realize multi-performance model matching. BTF (Brightness Transfer Function) is also used to eliminate the color feature differences between multiple cameras. Experiment show that BTF eliminates the differences efficiently and improves the matching accuracy.Thirdly, considering the topology relation of non-overlapping multi-camera, Gauss model is used to describe the probability distribution of transfer time, this provide object matching a time clues.Finally, in order to sufficiently use the various informations and improves the accuracy of object associating, we adopt the weighting sum to merge the multi-performance and the probability distribution of transfer time.