Filtering in a unit quaternion space for model-based object tracking

The main idea in object tracking is to support the processing of incoming images by predicting future object’s poses and features using the knowledge about the object’s previous motion. In this paper we present a new method for the prediction and adjustment of motion parameters to the current measurements using the quaternion representation for the orientation and the Gauss-Newton iteration on the unit sphere S3. Unlike other trackers, our tracker searches for estimates directly in the space of rigid body motions (represented by R3 × S3) and takes into account the differential and geometric properties of this space. We present some results showing the successful tracking in synthetic and real image sequences.