Dense motion segmentation of articulated objects in crowds

This paper introduces an algorithm for dense motion segmentation of pedestrians in crowded video sequences. This algorithm realizes dense and temporally-consistent segmentation under severe occlusion conditions. To segment the whole appearance of each articulated object, the algorithm uses temporal invariance of geodesic distance (similarity) between segments as a criterion for motion segmentation. Furthermore, a random walk framework is applied to Intersegmental Spatio-Temporal Connectivity (ISTC) to obtain dense and temporally-consistent segments from short-term trajectories. The experimental results show that the proposed algorithm can obtain segments of 66 % density in average even in crowded scenes.