Research On Human Action Recognition Algorithm Based On 3D Data

Human action recognition based on computer vision has been widely used in many fields of our daily life.In recent years,with the release of low-cost deep cameras,people can easily acquire RGB and depth data of scenes.We also can extract the position information of human skeleton and joints from depth data in real time,which greatly simplifies the task of object segmentation and helps research and application of human action recognition.In this paper,we use skeleton features to recognize human action,and propose an algorithm that can recognize the human action in real time.On the other hand,we also integrates skeleton features and depth features to realize the recognition of complex interactive actions.The main research contents are as follows:Firstly,we study the method of selecting skeleton joints,and select partial joints that play a key role in the actions to compose the joint subsets.Moreover,we calculate sets of spatial and temporal local features from subgroups of joints.Secondly,the Hidden Markov Model(HMM)human action recognition algorithm based on gesture selection was proposed.This method uses two-layers affinity propagation(AP)clustering algorithms to select the key pose for each action automatically,which corresponds to the hidden state of the HMM.We use these hidden state tags to initialize the HMM parameters.The trained model replied on initial parameters is used to perform the classification.Compared with the traditional HMM algorithm,this algorithm can reduce the impact of initialization parameters on recognition performance.Compared with other recognition algorithms,it is more adaptable under different conditions and can be realized in real time.We apply the traditional K-means clustering algorithm and random initialize parameters on the MSR Action 3D database.Compared with our method,the result demonstrates the rationality of our initialization method.Compared with the domestic and foreign methods,the experiment result proves the effectiveness and real-time performance of our algorithm.The experimental results on the MSR Daily Activity 3D show that the algorithm has strong adaptability.Thirdly,in order to identify more complex interactions between human and object,we fuse skeleton features with the depth features around joints.The skeletal features cannot provide feature information outside the human body.But the depth features can capture the detailed features of objects.To handle the temporal misalignment and noise issues,we propose Fourier Temporal Pyramid to remove the high frequency coefficients.The extracted low frequency coefficients increase temporal dynamic characteristics of these extracted features.The Canonical Correlation Analysis Algorithm(CCA)is used to fuse the two kinds of features.Action recognition is performed by classifying the final feature vectors using one-vsall linear SVM.We also verifies the recognition rate under different layers and select the most suitable three pyramid layer.The final results on the MSR Daily Activity 3D database demonstrate the effectiveness of the feature fusion algorithm.