Making tracking of object in image sequences steady despite complex movement of objects

This paper begins with an analysis of the shortcomings of existing mean shift algorithm. We aim to overcome these shortcomings with our improved mean shift algorithm in which we introduce two distinguishing features: the bandwidth matrix and the target angle. We first introduce the bandwidth matrix mean shift procedure. Then we describe the target by introducing the target rectangle, which provides two position coordinates of the center point, the horizontal axis, the vertical axis and the target angle, altogether five degrees of freedom. Target angle is used to accommodate the rotation of objects (this accommodation was not provided by past researchers) while the two axes determine the size in two independent directions. The dissimilarity between the target color distribution and the target candidates is expressed by the metric derived from the Bhattacharyya coefficients. Experiments with several real world's sequences indicate our new method's capability to adapt to any combinations of the target rotation, zooming and translation. With better description of the object it achieves much better precision. The comparison of computational burden with two existing methods demonstrates the possibility to apply our algorithm in real time.