Motion segmentation by velocity clustering with estimation of subspace dimension

Abstract

The performance of clustering based motion segmentation methods depends on the dimension of the subspace where the point trajectories are projected. This paper presents a strategy for estimating the best subspace dimension using a novel clustering error measure. For each obtained segmentation, the proposed measure estimates the average least square error between the point trajectories and synthetic trajectories generated based on the motion models from the segmentation. The second contribution of this paper is the use of the velocity vector instead of the traditional trajectory vector for segmentation. The evaluation on the Hopkins 155 video benchmark database shows that the proposed method is competitive with current state-of-the-art methods both in terms of overall performance and computational speed. © 2013 Springer-Verlag.