Applications Of Three-dimensional Signal Processing For Object-tracking And Image Compression

In this contribution, two distinct methods are described for the processing of 3D image sequences. The first method is voxel-based and employs narrow band 3 0 recursive filters, directly in fhe spafio-femporal (Sr) domain, to extract visually acceptable information. The filters ideally have support (i.e. a 3D passband) that closely surrounds a 3D plane in the 3D frequency space. Some object enhancement and tracking applications are demonstrated. The second method is object-based and is useful for high compression of talking heads for videoconferencing applications. A 3D spatial model of the dynamic human head is described that employs temporal actuation of muscles to deform the surface of the 3D facial model in order to create realistic facial expressions. The resultant compression of data requires the a priori availability of the sfafic 3D head model. Realistic dynamic facial expressions may be transmitted in the form of muscle actuation and head positional parameters. The potential exists for high compression of talking head data and for operation at real-time video rates. Other advantages of 3D model-based processing of heads are that the viewer may occupy an arbitrary position in 3D space from which to view the original scene and the reconstructed 3D model is voxel-independent and therefore independent of the resolution of the display device. Further, the viewer may create a visually meaningful simulated environment in which speech recognition techniques may be coupled to the model to achieve lip synchronization.