Discovering And Exploiting The Structural Pattern Of Feature Vector Generated By Bag Of Visual Words Model Based Methods For Human Action Recognition

Human action recognition is an important research area in computer vision,which has been widely applied for video surveillance,video retrieval and human-computer interaction,etc.The bag-of-visual-words?BOVW?model based human action recognition methods usually generates high dimensional feature vector,which is not good for the improvement on classification accuracy.To address this,the thesis discovers the structural pattern hidden in the high dimensional feature vector generated by BOVW-based methods,and regards the structural pattern as prior information for the classification process,then embedded it into classifier by structured regularizer for better classification accuracy.The contributions of the thesis are summarized as the following:Firstly,the discovery and exploitation of the structural pattern hidden in the feature vector generated by the“hard assignment”method is studied.The“hard assignment”method is the most common BOVW-based method,this thesis clusters the visual words and regards the clustered results as group structure of feature vector,and proposes a group sparsity regularized support vector machine?GSRSVM?for human action recognition by substituting group sparsity regularizer for l₂ regularizer in SVM.An algorithm based on alternating directions method of multipliers?ADMM?is presented to optimize the GSRSVM.Experimental results of human action recognition task demonstrate the effectiveness of the proposed method.Secondly,the discovery and exploitation of the structural pattern hidden in the feature vector generated by the Fisher Kernel method is studied.This thesis finds group structure of feature vector generated by Fisher kernel,and regards the variables describing the same Gaussian component as the same group,then exploits the group structural information via GSRSVM.An optimization algorithm based on dual function for GSRSVM is introduced,which enables the GLRSVM to handle extremely high dimensional classification task.Experimental results of human action recognition task demonstrate the effectiveness of the proposed method.Thirdly,a sparse coding combining spatio-temporal pyramid method is presented,and the discovery and exploitation of the structural pattern hidden in the feature vector generated by the method is studied.The presented method has not only taken advantage of the spatio-temporal structural information about visual words,but also reduced the quantization error.More importantly,there are two kinds of structures hidden in feature vector generated by the proposed method,which are group structure and hierarchical structure.The group structure information is exploited by GSRSVM.Whereas for the exploitation of hierarchical structural information,a hierarchical sparsity regularized support vector machine?HSRSVM?is proposed by substituting hierarchical sparsity regularizer for l₂ regularizer in SVM.An algorithm based on ADMM is proposed to optimize the HSRSVM.Experimental results of human action recognition task demonstrate the effectiveness of the proposed method.Fourthly,a multi-level sparse coding method is presented,and the discovery and exploitation of the structural pattern hidden in the feature vector generated by the method is studied.The proposed method has not only prevented the propagation of the quantization error,but also captured meaningful contextual information.What is more,the feature vector generated by the proposed method has a group structural pattern,and embeds it into GSRSVM.Experimental results of human action recognition task demonstrate the effectiveness of the proposed method.