Currently, we are interested in using active snakes as a general framework for accurate boundary localisation and tracking within a studio environment. The problem domain requires a precise segmentation of moving non-rigid objects, e.g. people, whilst circumventing the necessity for preliminary environmental preparation. By obtaining a thorough experimental understanding of the active contour paradigm within such a domain, the theoretical and experimental issues for research are to be discovered.
At present, we have combined a translational motion estimation framework with local gradient optimisation to track objects with (planar) movement in the image plane. The translation parameter (image displacement) is recovered using intensity profile matching, based on recent work by Rowe & Blake (ECCV'96). This essentially transforms contour attachment into a local deformation problem. Other work has also been done using so-called "natural basis functions" (Rao & Ballard, ICCV 1995) for recovering translation and scale parameters. Plans for the near future involve increasing the robustness of angle parameter recovery through a number of linked NBF foveal areas, interpolation between responses of different scales to recover scale parameters, and point distribution models for refining snake attachment.