Human Behavior Recognition Based On Human Skeleton Model

Recently,human body recognition becomes a very popular research topic.It can be applied in video retrival,intelligent retailing,human-computer interaction,public safety,and action generation.Therefore,it attracted the attention of the academia,the industry and,even the government.But human action recognition based on video is still a challenging problem due to the complexity of human action,the definition of basic action,the view of photograph,the illumination condition,and the background and the other cause.In this thesis,we concentrate on the key problem in human action recognition including the feature used in action recognition,human skeleton feature,human body model,parameter optimization,action description based on human skeleton feature.The main contents of this thesis are showing as follow:1.Analysing three human body model named volumetric model,cardboard model and matches model.We choose matches model as our human body model and modify it to adapt to our dataset with the criterion 'without loss of generality'.HOG descriptor and some geometric features are chosen as basic features for building human body model.then,a tree-structual graph is used to descript the human body model and the objective function of optimization the parameter of human body model is proposed.2.Comparing the algorithms that can solve the objective function including gradientbased methods,SDCA-like methods.We introduce 'momentum' to the proximal SDCA which can get a accurate solution and observe whether the 'momentum' can accelerate the proximal SDCA.Then,we use the Nestrove's estimate technique to compute the coefficient in the accelerating algorithm automatically,which improves the effectiveness of the accelerating algorithm.3.Three methods are used to descript the video with human action.First,FFT is used to transform the feature in time domain to frequency domain which makes the feature aligned.Then,Clustering human pose according to its status,and encode the feature being clustered in codebook.Finally,Extracting the trajectory and HOG features around the trajectory and modeling those feature with GMM,followed by encoding those feature vectors with fisher vector.4.To testify the practicality we conduct the experiments about the parameter optimization among diffrent algorithm mentioned in this thesis and the description of action.Experimental results show that the effectiveness of parameter optimization is significantly reduced and the action recognition also has a good result.