Object tracking algorithm of adaptive kernel-bandwidth for mean-shift based on optical-flows

To obtain and update the kernel-bandwidth, we present an adaptive bandwidth obtainment algorithm based on object contour extraction from optical-flow field. The combination of modified mountain cluster approach and fast scanning window contour extractor guarantees the speed of this algorithm. A novel ellipse detection method based on a modified RANSAC is adopted to reduce the noise. Experimental results demonstrate that the algorithm select the proper size of tracking kernel-bandwidth with minor extra computational overhead and keep up with the object robustly when the scale changed rapidly.