Research On Human Motion Learning Method Based On Probability Model

In recent years,with the rapid development of computer vision technology and depth cameras,research on reliable segmentation and recognition of human poses based on human poses and other visual information has become more and more in-depth.Human gesture recognition allows robots to recognize human movement characteristics by combining environment and purpose,and learn human actions,so as to have the ability to learn the complex behavior of humans,and even interact with humans in a dynamic environment.It is one of the hot spots in the field of intelligent robot research.From the perspective of motion learning,the process of motion recognition can be seen as an understanding of continuous human behavior,which is composed of a series of actions,such as waving,picking up,putting down,and so on.The key issues are how to identify what actions are in a given domain,how to encode them,and how to identify them in the captured motion data.This paper analyzes the main research methods and research status of motion learning.On this basis,based on the Gaussian Mixed Hidden Markov Model(GMM-HMM),the human motion learning method of humanoid robot is studied.First of all,the skeleton and joint model extracted by Kinect is preprocessed,and the 3D coordinate sequence of skeleton and joint is transformed into the joint angle feature sequence by space vector method,and use the quaternion to represent the skeleton pose feature sequence,and the data sequence is normalized at the same time.Secondly,group training and testing of the GMM-HMM is performed on the authoritative database MSR Action3 D to determine the optimal parameters of the model;In order to improve the stability of the model,in the grouping of data sets,the action performers of each training set and test set are different from those of other groups,and a half-and-cross grouping method is used which is different from the conventional one;in addition,the recognition effects of coordinate data sequences and joint angle feature sequences are compared and analyzed.Finally,a skeletal posture model represented by quaternion is established,and it is proved that it is superior to the original coordinate data and the recognition effect of the joint angle feature sequence;After that,a skeleton pose model combined with Kinect was used to identify human movements in real time,and an improved limiting filtering algorithm was proposed based on the confidence value in the acquired data.The experiment verified the stability and real-time performance of the method used.