Tracking with generalized active contour models

Abstract

This thesis considers the problem of motion tracking and analysis of deformable contours based on generalized active contour model (g-snake). We propose a framework which encodes specific knowledge on the shape, motion and deformation of the tracked features. Using these information, the trackers perform contour synthesis, localization, refinement and match operations. We suggest four trackers and confirm their validity through extensive experimentations. The first tracker overlays the preceeding g-snake on the new image frame, and then restarts contours refinement to obtain the best match template. In order to exploit temporal redundancy existing in image sequences, the second tracker imposes motion smoothness constraint to perform adaptive motion prediction. The third tracker applies principal component analysis to synthesize a codebook of contour templates. By combining these ideas, the last tracker synthesizes templates along the major modes of deformation. Since these trackers, with the exception of the first tracker, require only a few parameters to describe the shape and motion changes of image features, they are suitable for very low bitrate image coding. We thus propose a model-based facial image coding framework in which g-snake trackers serve as a main component.