Mean shift based algorithm for tracking objects with changing orientation

In standard mean shift tracker (SMST), the object position can be well located, whereas its orientation cannot be determined. In this paper, a modified mean shift tracker is proposed to obtain the object position and orientation simultaneously. By defining the concepts of feature angle and position angle, a new object model is formulated, where the feature space, originally only containing the color feature in SMST, is augmented by adding the angle features of pixels, and the weights of pixels depend not only on its normalized distance from the object center but also on its position angle. Based on the new object model and the Bhattacharyya similarity function, object tracking is achieved by solving an optimization problem where the cost function depends on both the position and the orientation of the object. Through alternate iterated optimization, both the position and the orientation of the object can be determined. Extensive experiments are performed to testify the proposed algorithm and validate its robustness to the orientation changes of the object.