Research On Sparse Subspace Clustering Algorithm And Its Application In Motion Segmentation

Now people are not just satisfied with playing multimedia information,but moving to the access,retrieval,and operation of video objects.So,video-based motion segmentation technology has become the focus in research.Motion segmentation can separate objects with different motions from others in the video,and it is the cornerstone of object-based video coding,video retrieval,and multimedia interaction.The traditional motion segmentation algorithms use the methods of moving target detection and target tracking.When using the frame difference method and the optical flow method to detect the moving objects,the results can be very susceptible to noises.Meanwhile,the targets occlusion,distortion,and deformation and some other problems are also involved in targets tracking.It is hard to get the idea results of motion segmentation in complex scenes.To avoid the problems encountered in motion detection and target tracking,this paper adopts sparse subspace clustering algorithm to realize motion segmentation in complex scenes.Because feature points which have the same motion trajectory are on the same linear manifold,so we can use the sparse subspace clustering algorithm to achieve the feature points motion segmentation.When processing high-dimensional data,sparse subspace clustering algorithm can segment these data into low-dimensional subspace.Meanwhile,the algorithm can deal with the impacts of singularity and noise.The following work is done for the study of sparse subspace algorithms.(1)By comparing k-means algorithm,adaptive spectral clustering algorithms have been studied deeply.Because sparse subspace clustering algorithm is based on spectral clustering,we have done an in-depth research of spectral clustering and analyzed the status quo of spectral clustering research and application.To overcome the shortcoming of spectral clustering,our algorithm can determine the number of clusters automatically based on the matrix of the perturbation theory,and calculating the characteristics of the matrix gap.To prove the spectral clustering algorithm can deal with arbitrarily shaped data set,and do not fall into the local optimal,this paper selected a variety of shapes of sample sets.Compared with k-means algorithm,the experimental results show that The Adaptive Spectral Clustering Algorithm performs better when dealing with these sample sets.(2)Hybrid least-squares regression sparse subspace clustering algorithm isproposed.To ensure the largest similarity of data in the same class and the smallest similarity of data belonging to different classes,the similarity matrix must be sparse and homogeneous.By analyzing and focusing on each data which indicates the maximum sparsity of the coefficients,we found that sparse subspace clustering lacks the description of the global structure of the data sets.The low rank subspace clustering algorithm guarantees the structural relevance of the same kind of data,but it is not sparse.In this paper,we use the data matrix to deal with the various noise points,singular sample points and isolated points in the sample sets.And we introduce the least squares regression into the sparse subspace clustering algorithm to ensure that the similarity matrix of the data has sparseness and grouping effect.The improved algorithm performance is validated by experiments.(3)Research on sparse subspace clustering algorithm in the application of motion segmentation.The sparse subspace clustering algorithm is applied to the video object processing,and the motion segmentation model is established.The motion segmentation experiment is carried out.The experimental results show that the proposed algorithm can improve the clustering accuracy of the algorithm in the case of time complexity.