Video Object Tracking Based On Temporal-spatial Constraints

Thanks to the dual driven of technological development and social needs, the videotracking is widely applied in security surveillance, military and behavior understanding. Alarge number of video object tracking methods have been proposed, while the study of along-term effective video tracking method is still an urgent problem because of theillumination changes, object occlusion and lack of an adaptive representation model to theobject changes. Although there are many challenge problems in video object tracking, it iswith important significance for its potential commercial value.The thesis proposes an improved Mean Shift tracking method based ontemporal-spatial constraints to handle the problem of scale changes and occlusion. Thetemporal and spatial constraints of the image sequences are proposed to improve thetracking performance. The extended kalman filter is exploited to predict the target positionas the starting point of the Mean Shift tracking. The thesis takes advantage of the object’smotion constraints and the offset difference of the forward-backward tracking is used tocompensate the scale changes of the object. Besides, the pictorial-structures are used toimprove the detection performance under cluttered, similar and occlusion environment.The specific content of the thesis are summarized as follows:(1) The thesis takes advantage of the object’s motion constraints; the extended kalmanfilter is combined with the Mean Shift tracking. The predicted result of the extendedkalman filter is utilized as the iterative starting point of the Mean Shift. The movementtrend of the object is considered to improve the robustness of Mean Shift. Theforward-backward tracking is used as the basis to adjust the size of the tracking boundingbox in a certain extent to follow the object scale change adaptively; while as the basis toimprove the tracking accuracy simultaneously.(2) For the complexity of video object tracking, an improved Mean Shift trackingmethod which is combined with detection is proposed based on temporal-spatialconstraints. When in the situation of severe interference, the cascade classification methodwhich consists of nearest neighbor and self-organizing maps classification is utilized todetermine the image patches which are the most likely to contain the object. The proposedmethod makes full use of the temporal and structural constraints to improve the trackingperformance. The detection module re-position the object to obtain a complete trajectorywhen object drift and lose are occurs.