A New Object Tracking Algorithm Based on Mean Shift in 4-D State Space and On-line Feature Selection

Visual tracking using Mean Shift is famous and popular. But the traditional Mean Shift tracking algorithm cannot track an object which changes its scale and orientation during the process of tracking. A novel tracking algorithm based on Mean Shift and on-line feature selection is proposed in this paper. Target object is defined in a 4-D state space which can deal with its positon, scale and orientation. A maximum of 28 feature space is created based on the color value of pixels in R, G, B channels. During the tracking, the best feature space is selected which can distinguish objects and background scenes most. Kalman filter is used to estimate the state of the object during the tracking. Experiment results show the advantages of the proposed algorithm.