Motion Segmentation Based On Subspace Clustering

Motion segmentation is the basis of the analysis and understanding of the scene video sequences. In the fields of computer vision and pattern recognition, motion segmentation is an important research content. It is widely used in automatic navigation, security monitoring, intelligent transportation, and so on. In recent years, with the swelling of the amount of information in the human world, and the arrival of the high-dimensional data, motion segmentation methods are also improving constantly and adapting to the new environment constantly. When faced with high dimensional data, noise interference, camera movement etc., traditional motion segmentation methods while have all sorts of problems, and it becomes very difficult to achieve the ideal segmentation results. In this background, motion segmentation method based on subspace clustering, quickly become one of the most effective way to deal with motion segmentation problem. Compared with traditional motion segmentation methods, methods based on subspace clustering have many advantages, especially methods based on spectral clustering such as the low rank representation subspace clustering and the sparse representation subspace clustering. These two methods compatible with high dimensional data, have good robustness for the problem such as noise interference, camera movement etc. Motion segmentation method based on subspace clustering firstly trying to find a similarity matrix which measured the similar degree between every two data points according to relevant optimization algorithms. The similarity matrix is described as that data points in the same subspace own the larger similarity measure, and data points in different subspace own the smaller similarity measure. We build a similarity graph based on the similarity matrix, data points correspond to graph nodes, similarity between data points correspond to edge weights. Finally, we cluster graph nodes to different subspace using relevant spectral clustering methods, then we realize the division of clusters, and complete the motion segmentation.This paper proposes a new motion segmentation method based on consistency sparse subspace clustering. Sparse subspace clustering extract low dimensional substructure in the high-dimensional data through seeking the sparse representation of data points about other data points, which did not consider the possibility of belonging to the same subspace for different feature points and ignored the motion consistency among them. In view of this problem, we define a "consistency measurement model" based on the consistency level of tracking points' behaviors in the video. This model accurately measured the possibility of data points belonging to the same subspace before solving the sparse representation of the data. According to this model, we propose the Consistency Sparse Subspace Clustering algorithm(CSSC) for motion segmentation. The proposed algorithm introduces the motion consistency to the objective function of the sparse optimization model, getting the sparse optimization model with the consistency constraint, and then resolves the new sparse optimization model using alternating direction multiplier method(ADMM). Under the minimization constraints, the new model is more advantageous to isolate big coefficients in the sparse representation, and the feature point trajectory is more advantageous to be represented as a linear combination of other feature point trajectories belonging to the same moving object, thus we can get the more optimal sparse representation coefficient. The experimental results show that the proposed model can get the more optimal sparse representation, and to a certain extent, the introduction of consistency constraints can take a part of feature points divided by mistake back to the cluster which own the consistent movement, making the CSSC algorithm be able to correctly identify a large number of feature points which belong to the background motion. The proposed algorithm has the better performance on the clustering accuracy comparing with other similar algorithms, especially for the division of feature points which belong to the background motion.