Motion Segmentation Based on Independent Subspace Analysis

In this paper, we propose a novel method to address the segmentation problem of multiple independently moving objects. Based on the fact that multiple objects’ trajectories correspond to multiple independent subspaces, first, bases of these subspaces are extracted by applying independent subspace analysis (ISA). Then, these bases are grouped properly after evaluating the correlation coefficients of them. Feature grouping and outlier rejection are effectively performed by calculating the data point’s membership functions to these subspaces. A reasonable energy function is also introduced to facilitate optimal segmentation. The geometrical essence of the method is regarded as a global constraint added in the segmentation process resulting in a considerable increase in error tolerance, without either prior knowledge of the number of objects or prior assumption about existence of degeneracy. The experimental results on synthetic and real data both demonstrate the effectiveness of our algorithm.