Motion Segmentation Based On Spectral Clusteringand Point Features

Motion segmentation is an important application in the field of computer vision and visual information processing.It is the basis of video scene analysis and target tracking technology,so it has been applied in many aspects,such as video surveillance,auto autonomous driving,automatic navigation,automatic tracking and other fields.Video data in the rapid expansion in recent years,the traditional clustering algorithm in the face of such a high dimensional data is very weak,so the improved algorithm is more and more attention has been,in the same way,in the field of motion segmentation,high-dimensional data,outlier data,view of uncertainty poses challenges to the traditional algorithm,Therefore,it is necessary to improve the existing motion segmentation algorithm.Recently,a new algorithm named subspace clustering has solved the problem of highdimensional data well.The method of subspace is used to reduce the dimensioning of highdimensional data,and the obtained data after dimensioning is used to construct the similarity matrix,and then K-means algorithm is used to process the clustering results.However,this method also has a lot of limitations.It does not perform well when dealing with video sequences with too many outliers,and it is not very stable even when facing videos with varying angles of view.This paper will be aimed at the above several issues on the research.When moving objects are segmented,there are always some problems such as the blocking of moving objects,which can not be solved by traditional motion segmentation methods.For Hopkins155 data set,the scenes of most videos are relatively small,so the most suitable matrix model should be used to fit the video motion to segment such videos.To solve these problems,a motion segmentation method based on similarity matrix and homography is studied in this paper.The homography matrix model is used to fit the scene.The similarity matrix method has a better effect on solving the occlusion problem.From the analysis of the experimental results,this method based on homography matrix has made a lot of progress in the accuracy of motion segmentation of Hopkins155 data set.However,the effect of homography matrix is not very good if it is changed to other data sets with large perspectives.Therefore,this paper also studies a method based on complementary geometric model,which integrates the advantages of affine matrix,homography matrix and fundamental matrix,so that the algorithm can also have a good effect when there is a real data set.Traditional algorithms pay too much attention to motion segmentation based on trajectory,resulting in failure to improve accuracy.However,trajectory data in the data set all need to be processed manually,so it is not stable to deal with other video data sets.Therefore,this paper also studies a motion segmentation method based on feature points and pair matching.Because it is a new idea,some experimental results are not as good as the previous methods,but it can be applied to ordinary data sets.The experimental results show that this method also achieves high accuracy.