Human Motion Recogntion Based On Three-dimensional Skeleton Model

Human action recognition refers to the pattern of human movement analysis andidentification. It is a hot research problem in computer vision and pattern recognition.Human action recognition is widely used in the intelligent monitoring, film actionsproduction, human-computer interaction, motion analysis, medical rehabilitation and otherrelated fields and also creates a lot of social and economic benefits. Based onunderstanding of the research status of human action recognition and analysing of relatedaction recognition algorithm, this paper studied the human skeleton model to expand thestudy of human action recognition. The main work is as follows:(1) We construct a new three-dimensional model of the human skeleton by analysingthe traditional mannequins. The influence of the distance from the person to the camera onthe human body model will be eliminated if the human body model is normalized. Thenwe establish the YsuMan3D human motion database by using Kinect. The databaseincludes walking, waving, boxing, answer calls and other daily actions.(2) For recognizing the single perspective human action, we proposed a recognitionmethod based on phase space tracking. Firstly, the motion characteristics are extractedfrom the human body skeleton model. Secondly, Takens’ theory is applied toreconstructing the phase space of the characteristic action sequences. Based on DynamicTime Warping (DTW) algorithm and the Hausdorff distance, we propose a new method ofcalculating the distance between phase portraits. We use the weighted method for trainingaction model and to achieve the action recognition. Finally, in order to evaluate ourmethod, we have conducted experiments by using two actions databases. Theexperimental results show that our method has a certain universality and robustness.(3) For multi-view human action, the method based on improved single-view actionrecognition is proposed. Firstly, we use respectively the least squares method (LS) andprincipal component analysis (PCA) to extract mannequin direction vector, and then buildthe local coordinate system of the human skeleton model. Then the support vectormachine (SVM) method is used to train the action model. Finally, the improvements of the method of the single-view action recognition are completed, and multi-view human actionrecognition is achieved. It is laid a solid foundation for further human action recognitionapplications.