Oriented particle filter for visual tracking

The problem of tracking general objects is still challenging. Particle filter is inadequate in many cases because the results often contain nothing more than the trajectories of moving objects and no contour information is involved. We propose a new method, the Oriented Particle Filter. Our method employs the Gaussian Mixture Model to represent the object, and incorporates spatial cues by assigning an orientation to each particle. Experiments show that the OPF is robust to partial occlusion and appearance change, more importantly, the tracking results are much more accurate in describing the contour changing of the object. The model updating mechanism and initialization technique based on semantic description are also discussed.