Target Tracking Algorithm Based On Emd And Implementation

Object tracking is one of the most important disciplines in computer vision and haswidely applications in the fields of civilian, military and engineering. This thesisprimarily studies object tracking based on EMD (Earth Mover's Distance).The thesis firstly proposes EMD based similarity and affine transform trackingmethod, it is an improved algorithm of DEMD (Differential EMD). The original DEMDalgorithm appearances object using a vertical ellipse, it can't tracking similarity oraffine object. The proposed tracking method addresses the issue of this problem, itfirstly modeling object using similarity and affine matrix, optimization the scorefunction, and then get the object contour. The algorithm is mainly applied to trackingsimple geometric deformation (such as translation, rotation, scale and shearing) objects.In order to track the object with complex transformations, the thesis proposes theEMD based PDE (partial differential equation) tracking method. The EMD based PDEtracking algorithm firstly modeling the object by region functional, then we derive thePDE that governs the contour evolution. Compared with the Bhattacharry distancebased PDE tracking method, the proposed method is insensitive to illumination changes,and allowing part of the match. In the thesis, we also compared the tracking resultsbetween two different features (SIFT and color) in the same tracking method.A tracking system must be able to tracking the target real-time and accurately.Although the tracking algorithm works well, but the tracking time is verytime-consuming. For decreasing the tracking time, a parallel PDE tracking method isproposed, and the parallel parts include computing EMD and PDE. The experimentshows that the tracking time is decreased significantly.