Adaptive bandwidth mean shift algorithm and object tracking

The classical mean shift algorithm is extended to be the adaptive bandwidth mean shift algorithm, and then the adaptive bandwidth mean shift object tracking algorithm (ABMSOT) is proposed. The former gives the general adaptive bandwidth mean shift framework for seeking the local maxima, and the latter can simultaneously tracks the position, scale and orientation in real time. For ABMSOT, the feature histogram weighted by a kernel with adaptive bandwidth is used to represent the target model and the candidate model. Similarity of the target model and the candidate model is measured by Bhattacharyya coefficient. A two step method is used iteratively to find the most probable target position, scale and orientation. The first step finds the object position using a mean shift iteration, and the second step finds the bandwidth matrix which best describes scale and orientation of the object region. The convergence of the two algorithms is proved theoretically. Experiments show that ABMSOT can successfully track the position, scale and orientation in real time.