Research On Action Recognition Based On Non-negative Matrix Factorization

In the era of massive data,how to effectively deal with large-scale data,such as videos,and how to mine the effective information,are problems to be urgently solved of the science and engineer.Non-negative matrix factorization?NMF?is an effective matrix factorization method.Under the non-negative constraints,this method can effectively realize the data dimensionality reduction,the factorization results of NMF conform to the intuitive experience of human cognitive process,and possess better interpretability.Thus,NMF has been widely used in many fields such as pattern recognition,feature extraction and signal processing.On the basis of deep analysis and research on the existing NMF methods,this thesis presents the following two kinds of improved NMF methods,and uses these methods for human action recognition.1.A Local Weighted Orthogonal Non-negative Matrix Factorization?LWONMF?method.Firstly,this thesis assigns the weights to the element of coefficient matrix by constructing the local weight operator,takes this as local constraint term and introduces it into the objective function.And then,to obtain the clear clustering structure,the orthogonal constraint is also incorporated into the basis matrix.Using these two constraint terms makes the LWONMF method possess the clear clustering structure and the intra-cluster structure of cluster more compact.Meanwhile,the update rules of basis and coefficient matrixes are derived by using the simple and effective multiplicative update algorithm.Experimental results show that,the convergence speed of LWONMF method is fast,and in the clustering contrast experiment,the LWONMF method can fit the motion saliency regions well,and also fit the discrete trajectories in the background.The motion salient regions obtained on the basis of the result can alleviate the passive influence of complex background on feature extraction.2.A Multilayer Non-negative Matrix Factorization?Multilayer NMF?method.Firstly,a Temporal Smooth Constraint Non-negative Matrix Factorization?TSCNMF?is proposed.In this method,considering the time-dependent characteristics of frames for video,the temporal dependencies constraint term is introduced into the objective function to extract the spatial-temporal feature of video.Furthermore,the sparse constraint term is also introduced to improve the representativeness of feature by using theL_2,1 norm.Meanwhile,the update rules are derived.With the help of hierarchical feature extraction strategy,Multilayer NMF is proposed by constructing the multilayer structure,which uses the TSCNMF as the unit algorithm.This method takes the constructed non-negative matrix as input,and decomposes it layer by layer,which refines the feature.Combining the results of each layer and enriching the feature representation can improve the representativeness of feature and are helpful for human action recognition.Furthermore,the experiment part gives several experiments that use the feature obtained by using Multilayer NMF for human action recognition,and the results shows the effectiveness of method for this thesis,which obtains good recognition performance.