The Research On The Kinect-based Action Recognition

Action recognition is one of the important research topics in computer field,which has a wide of applications such as smart home,health monitoring,security and remote immersive applications.With the appearance of Microsoft RGB-D camera i.e.Kinect,it has gradually revealed its powerful function in the field of action recognition,which provides a new way for the action recognition as it has the advantages such as low computation,high robustness and strong anti-interference ability.Kincet can provide 3D information about human bone nodes,which provides high dimension data to further improve the accuracy and the recognition efficiency.However,there are still some problem such as low accuracy and high compuation cost which constrains the further applications of Kinectbased action recognition.Herein,this thesis will focus on these problems.To combat the problem of low accuracy and high computation cost due to the skeletal representation,a bone hierarchy model is designed in this thesis.It divides the human skeleton data into six layers,and removes the redundant information by using the relationship between layers.Then,the covariance is calculated to encode the skeletal data,and the SVM classifier is used for activity classification.Experimental results on two databases show that this method can improve the recognition accuracy about 3% and the efficiency 9.375% compared with the current research results.To combat the problem of low recognition rate due to the skeletal coding for the local movement,this thesis integrates the scheme of joint motion simility into the above model.It divides the human body into several clusters on the hierarchical model.And multiple one-vs-all SVM classifiers are used for these clusters.Then the logistic regression is used to assign weight for each cluster,and the weighted distance of the sample distance from the hyperplane is obtained for the classification.Experiment results show that the recognition rate increases about 2.97% compared to the above model on two classic subsets.