A Region-Based Video Coder Using a Forward Tracking Active Mesh

Abstract

In an earlier study, we have reported the use of an active mesh representation for describing a video sequence [1]. In this representation, a scene adaptive mesh is generated for an initial frame, such that, the nodes fall on distinct features in the image and each element covers a 2D patch corresponding to a single object. The resulting nodes are then tracked in successive frames, so that the color patterns over corresponding elements in these frames match one another. The motion field between every two frames is interpolated from the displacements of the control nodes. In this paper, we present a coder that employs the active mesh structure. With this coder, an image is partitioned into regions of distinct motions. Each region is described by an ensemble of connected quadrilateral elements embedded in a mesh structure. The shape and color of each region are described by the nodal positions and color functions of the elements in this region in an initial frame; while its motion (including shape deformation) is described by the nodal trajectories in the following frames, which are in turn specified by a few motion parameters. Each motion class in this coder corresponds to a region (not necessarily connected), which may or may not correspond to a separate physical object. For this reason, we call this coder a region-based coder. This coder has been applied to a typical CIF resolution, head-and-shoulder type of sequence. At 50 Kbits/sec (for the luminance component only), the algorithm can render the motion more accurately and naturally than the H263-TMN4 algorithm.