Research And Development Of Action Recognition System Based On 3D Human Body Skeleton

Human action recognition is one of the important research contents of computer vision intelligence and pattern recognition.It has been widely used in many fields such as intelligent monitoring,natural human-computer interaction,physical and sports analysis.This paper takes the 3D human model collected by Kinect as the study object and carries out the research work on its action recognition.The results obtained are as follows:(1)A key feature construction method based on human skeleton model was designed.First of all,by integrating the actions of three common databases,the human action features are analyzed in detail on the basis of three-dimensional human skeleton model.Then,based on the spatial continuity of the motion,the two key features of the space relative position and joint Angle are extracted,combined with the change of bone point in the action sequence.It lays the foundation for the later action recognition.(2)In this paper,a Hidden Markov Model action recognition method is proposed,and the whole process of the Hidden Markov Model is analyzed.First of all,for the non-model method of feature extraction,only the defects of the lack of analysis of the human body structure are selected.Based on the model method,two key features,spatial relative position and joint angle,are used to obtain the feature vector that can fully express the human motion process.Then the action recognition experiments were carried out with MSR-Action 3D,Florence-Action 3D,UTKinect-Action 3D database and self-built database.Among them,by cross subject test,the given spatial relative position and joint angle combination characteristics were found that The average recognition rate of motion was 94.6%,95.7%,97.5%,and 89.2%,respectively,which was 8.6%higher than the average of the feature recognition rate alone,which validated the effectiveness of the action recognition method.(3)A real-time recognition method based on basic action is implemented.Firstly,combining with the early human skeleton model,the Kinect can get deep images and human skeleton data in real time,and the human bone points are integrated into the deep images.Then,the defined of 9 groups common human actions are recognized in real time.The experimental results show that the method presented in this paper is of real time and robustness.(4)A practical system for human action recognition is developed.Based on the method of this paper,a software system platform for action recognition is designed.The system can not only complete the training and acquisition of the action classification model,but also can use the Kinect to accurately recognize 13 kinds of actions,which has a good practicability.