Research On The Improved Trackin Galgorithm Of Mean Shift Using Joint Histogram And Anisotropic Kernel Function Respectively

Extracting the feature reasonably and rapidly, as well as real time locating the objectaccurately are the key points of object tracking. In the real world the scenes of objecttracking are complex and variable. Illumination variation and occlusion often occur. Thechanges of object itself such as scale changing, rotation and appearance change also bringdifficulties to tracking. Among lots of tracking algorithms, Mean Shift with advantages ofsimple calculation and stability become mainstream. In this paper we conduct deep study inMean Shift. And then propose improved methods aiming at the problems of tracking incomplex scenes. It mainly researches as follows:The single orientation of texture feature gets ineffective when tracking rotationalobject. In order to adapt to the complex-scene tracking, we present a kernel based trackingalgorithm which utilizes the joint histogram of colors and oriented gradients. The jointhistogram is made up of fusing colors and oriented gradients by expanding the dimensionsof the histogram. In the joint histogram principal feature is selected based on illuminationvariation factor (IVF for brief), which detects the degree of illumination variation. If IVF isbelow a certain threshold, color feature is taken as the principal feature. And then thealgorithm will be robust to rotation. Otherwise it chooses oriented gradients as principalfeature for its robustness to illumination variation. The problem of partial occlusion can befigured out by dividing the object template into sub areas. The presented algorithm showsgood performance in the experiments when dealing with complex scenes such as targetrotation, illumination change and partial occlusion.As the scale of classical Mean Shift is constant and the object template drifts, ananisotropic kernel function based tracking algorithm with adaptive scale and orientation ispresented. First Sobel operator detects the contour of object. Then the edge points that standfor the edge is extracted in uniform distribution direction. The kernel weights increase bydegrees from the edge to centroid. Then the anisotropic kernel function is constructed bycalculating weight refer to the edge points. Fixed incremental method is adopted to updateorientation information by choosing the angle that corresponds to max matching similarity.After locating object, the scale updates by strategy of equal proportion zooming. Then thetemplate is updated accordingly. The experiment results verify that the improved algorithm has better performance in the scene of object rotation and scale variation. So it deservesbroad application prospect.