New mean shift based algorithm for tracking targets with three degrees of freedom

Standard Mean Shift tracker can only successfully locate the object center, but fail to find its orientation, which make it not robust to track thin object. To remedy this, an improved mean shift tracker is proposed in this paper. The new tracker use new object representation, where pixels are weighted with both their position-angles and normalized distances from target center, furthermore, pixel's feature-angle, which can be seen as new feature, is introduced in. The new object representation can be conveniently integrated into the optimization framework of mean shift. By iterative optimization, both the location and orientation of targets can be precisely determined. Experimental results show the algorithm can get precise tracking results with low computational cost.