Segmentation-based object tracking using image warping and Kalman filtering

We propose a segmentation-based method of object tracking using image warping and Kalman filtering. The object region is defined to include a group of patches, which are obtained by a watershed algorithm. In a robust M-estimator framework, we estimate dominant motion of the object region. A linear Kalman filter is employed to predict the estimated affine motion parameters based on a second order kinematic model. Image (affine) warping is performed to predict the object region in the next frame. Warping error of each watershed segment (patch) and its rate of overlapping with the predicted region are utilized for classification of watershed segments near the object border. Applications of head and hand tracking using this method demonstrate its performance.