Evolving image segmentations for the analysis of video sequences

This paper concerns the segmentation of successive frames of a video sequence. Traditional methods, treating each frame in isolation, are computationally expensive, ignore potentially useful information derived from previous frames, and can lead to instabilities over the sequence. The approach developed here, based on the Region Competition algorithm (Zhu and Yuille, IEEE Trans. PAMI, 1996), employs a mesh of active contour primitives supervised by an MDL energy criterion. Temporal extensions, namely Boundary Momentum, Region Memory, and Optical Boundary Flow, are developed to ease the transition between successive frames. Further enhancements are made by incorporating mechanisms to accommodate the topological discontinuities that can arise during the sequence. The algorithm is demonstrated using a number of synthetic and real video sequences and is shown to provide an efficient method of segmentation which encourages stability across frames and preserves the quality of the original segmentation over the sequence.