Motion segmentation incorporating active contours for spatial coherence.

摘要

This thesis describes a computer vision algorithm that detects and segments independently moving objects in a video sequence, recovering their shape over time. While traditional motion segmentation approaches employ learned or low-dimensional parametric models to represent object shape, we propose a hybrid framework that combines robust motion segmentation with active-contour-based boundary recovery techniques, to overcome each individual approach's limitations. Our framework proposes feeding forward motion segmentation results to initialize, constrain and propagate the active contour, while feeding back active-contour-based object boundary estimates to the motion segmentation process to provide spatial coherence. We develop a functional system based on this framework, introducing a novel motion-based intensity constraint, and an active contour formulation that incorporates motion segmentation results. Our results demonstrate the successful segmentation of sequences that include multiple moving objects and sequences with a moving background.