Multi-feature Visual Tracking Using Adaptive Unscented Kalman Filtering

Visual tracking is often confronted with some impediments, such as the target´s sudden acceleration and structural deformation, occlusion, lighting changes and so on. To overcome these problems, a tracking approach is proposed, which is based on the unscented Kalman filter (UKF) and the multi-feature fusion. First, the mean and covariance of the target state variable is predicted based on a nearly constant velocity system. And the target´s hue histogram and edge orientation histogram are extracted at the corresponding position. Second, the measured position is calculated by Mean-shift algorithm based on the fusion of multi-feature. Finally, according to the measured position the UKF updates the mean and covariance of the state variable and reports the current position of the target. The experiments in 2 different scenes showed that the tracking method could efficiently track the fast moving objects and adapt to the lighting changes, rotation, and partial occlusion and deform. These demonstrated that the method have more tracking accuracy and adaptive robustness.