Improved motion segmentation using Locally sampled Subspaces

Motion segmentation is an important component of various video processing applications. In this paper an effective method for motion segmentation is presented. The method adopts the affine camera model. Initially, a local algorithm is applied to sample 4-subsets from the available trajectories. The Ordered Residual Kernel is then employed to measure similarities between trajectories. The algorithm proceeds by applying FastMap on the computed kernel matrix as a dimensionality reduction technique. The embedded vectors are used to produce an affinity matrix. Finally, spectral clustering is performed on the computed affinity matrix. Experiments on the Hopkins155 database demonstrate the robustness of the method to noise and its efficacy compared to existing approaches.