Object tracking based on multi-bandwidth mean shift with convergence acceleration

Abstract

A multi-bandwidth based tracking algorithm was proposed to search for the global kernel mode when the probability density has multiple peak modes. Firstly, a monotonically decreasing sequence of bandwidths was fixed according to the target scale. At each bandwidth, using mean shift to find out the maximum probability, and starting the next iteration at the previous convergence location. Finally, the best optimal mode could be obtained at the last bandwidth. To accelerate the convergence, over-relaxed strategy was introduced to enlarge the step size. Under the convergence rule, the learning rate was adaptively adjusted by Bhattacharyya coefficients of consecutive iteration convergence. The experimental results show that the proposed multi-bandwidth mean shift tracker is robust in high-speed object tracking, and perform well in occlusions. The adaptive over-relaxed strategy is effective to lower the convergence iterations by enlarging the step size.