Extraction and tracking of surfaces in range image sequences

Traditionally, feature extraction and correspondence determination are handled separately in motion analysis of (range) image sequences. The correspondence determination methods have typically an exponential computational complexity. We introduce a novel framework of motion analysis that unifies feature extraction and correspondence determination in a single process. Under the basic assumption of a small relative motion between the camera and the scene, feature extraction is solved by refining the segmentation result of the previous frame. In this way, correspondence information becomes directly available as a by-product of the feature extraction process. Due to the coupled processing of frames, we also force some degree of segmentation stability. First results on real range image sequences have demonstrated the potential of our approach.