The Analysis Algorithms For Human Motion Data

In recent years,with the rapid development of motion capture technology,the research of human motion based on data capture has attracted more and more attention.The progress of motion capture makes it easy for people to get a large amount of high quality capture data,which can be restored to the movement of the human body on one hand;on the other hand,the study of the data of these human movements can help people in virtual reality,medical rehabilitation,game scenes,military exercises and film and television.Production and many other fields have made greater room for development.The research of this thesis mainly includes two major problems: human motion assessment based on captured data and segmentation of human actions.The human action assessment based on capture data refers to the similarity between the capture data and a standard action,which includes both the similarity of the overall macro movement and the similarity of the local limb movements.The purpose of action evaluation is to give the similarity between two actions,especially the similarity between an arbitrary action and a standard action.In addition,motion evaluation can also provide action similarity measure for motion classification algorithm.The problem of human motion segmentation based on captured data refers to the accurate extraction of several independent semantic meaning movements from the continuous human motion capture data,such as hitting,jumping,running and so on.Effective segmentation algorithm can help to capture and reuse data.In the part of human action assessment based on capture data,we propose a G RU based human action classification model,which improves the accuracy,efficiency and generalization ability of the model in three aspects of data,features and models based on the LSTM based human action classification model.In terms of data,a variety of data enhancement mechanisms are used to expand the original data and increase the number of learning samples.In terms of features,multiple networks are used to extract features from various backbone departments of the human body instead of the original feature extraction for all joint single models.In addition to evaluating the similarity of the whole movement,it can also assess the similarity of specific parts of the body.In terms of models,GRU instead of LSTM greatly reduces training time and does not affect the accuracy of prediction results.The improvement of these three aspects can improve the fitting ability of the human action classification model and improve the prediction ability of the unknown action sequence samples,and also reduce the training and prediction time of the classification model.We propose two schemes based on the segmentation of the captured data,one is to determine the segmentation points on the basis of the classification model,to classify the action within a certain period of time and to change the action category.Then,it is divided into two ways: one way is a simple sliding window,setting the size of the window,moving the window forward with a specific step,and the other way is the proposed window,which can determine the distance of the next window to move forward according to the characteristics of the current window.It is the use of CTC.The idea divides the segmentation problem into a series annotation problem.CTC can automatically determine the segmentation points,eliminating the human annotation of the samples.We use the human action capture database HDM05 to experiment to verify the two classification methods we use.On the assessment classification problem,the classification accuracy of 98.94% and 95.98% is reached on the training set and the test set respectively.In motion segmentation,we use frame based accuracy,and the best average segmentation accuracy is 69.41%.