| Non-negative matrix factorization is an ideal dimension reduction algorithm.It requires that the elements of the factor matrix must be non-negative,that is,all elements must be equal to or greater than zero.This non-negative constraint results in the NMF being a part-based representation,which consistent with the cognitive process of the human brain.Therefore,it is an effective learning technique for image processing,processing and clustering.In this thesis,through the deep study of the existing NMF methods,two improved NMF methods are proposed and applied to the study of video human action recognition.(1)A non-negative matrix factorization with rank regularization and local weighted constraint(NMF_RRLWC)method is proposed.This method constructs two constraints,a local weighted constraint term and a rank regularization term,and applies them to the trajectory clustering.Local weighted constraints construct a weighted operator based on the similarity of the distance between the original data sample and the cluster center.The rank regularization constraint ensures that the decomposition result has a certain sparseness,and at the same time,it does not affect the manifold structure of the data due to too strict sparseness.Experimental results show that the addition of these two constraints can obtain better clustering results.The update rules and convergence proofs of this method are given in this thesis.In the experiment of trajectory clustering,the NMF_RRLWC can well fit the area where the motion subject is located,thus ensuring that the trajectories located in the background can be eliminated.(2)A Temporal Smoothness and Rank Regularization Constraint Non-negative Matrix Factorization(TSRRC_NMF)method is proposed for temporal subspace clustering.This method constructs a temporal smoothness constraint that takes into account the neighbor information in the time series data.In addition,in order to maintain the sparseness of the factorization results to extract the discriminative features,the rank regularization sparse constraint are introduced in the objective function.The update rules and convergence proofs of this method are given in this thesis.In the video action segmentation experiment,this method can obtain better video action segmentation result,which can correctly recover the temporal subspace structure of video data,so as to obtain relatively clear simple atomicactions.Based on this,a new feature extraction method for complex action video is proposed,which is a new feature representation for complex action recognition using simple atomic actions.A complex action usually exhibits richer temporal structures,which consists of a sequence of simple atomic actions.Therefore,the new method presented in this thesis transfers the knowledge learned from rich simple actions with labels to the recognition of complex actions,thereby improving the performance of complex human action recognition.The experimental results demonstrate the effectiveness of the proposed new method for the recognition of complex actions. |
