Robust motion segmentation via refined sparse subspace clustering

In this paper, a new refined sparse subspace clustering (RSSC) method is proposed for robust motion segmentation. Given a set of trajectories of tracked feature points from multiple moving object, RSSC aims at seeking a sparse representation (SR) for each trajectory with respect to a recovered low-rank dictionary. The segmentation of motion is obtained by applying spectral clustering to the affinity matrix built by this SR. Compared to the conventional sparse subspace clustering (SSC) algorithm, our RSSC integrates sparse representation and low-rank subspace structures recovery into a unified framework. Furthermore, SR is obtained from the recovered dictionary instead of the initial given dictionary built by contaminated data, making RSSC more robust to data noise. Experiments on toydata and real video sequences (Hopkins 155 database) show the superiority of our approach over several current state of the art methods.