Research On Human Action Recognition Based On Lie Group Features And Deep Learning

In the field of computer vision,artificial intelligence and pattern recognition,human action recognition has been attracting extensive attention,and its applications involve video surveillance,human-computer interaction,and medical rehabilitation,etc.However,due to the existence of the variability of human actions,temporal-spatial complexity and other problems,human action recognition still remains a challenging task.Based on the human skeleton information,this thesis proposes a method of human action recognition using Lie group features and deep learning.The human skeleton information is mainly used to overcome the interference of external factors such as changes of lighting conditions and body shape,the Lie group which belongs to popular structure is used to deal with the diversity and complexity of human motion data,and deep learning is used for high-dimensional data processing,feature learning and feature classification.Particularly,this thesis uses the Lie group skeletal representations to model the human action skeletal data,and the action features are learned and classified by CNNs.In the last part of the thesis,the application of human action recognition in virtual learning environment is studied.Following is the main research contents of this thesis:Firstly,the method of collecting the skeleton information of human actions based on Kinect is studied.This thesis briefly analyzes the basic working principle of Kinect,the process of collecting human action data and the method of human skeleton information acquisition.Secondly,the method of human action features extraction using Lie group skeletal representations model is studied.Based on the skeleton information,the human actions are modeled by using the relative 3D geometric relationship between various body parts,and human actions are modeled as curves in the Lie group.Then,the curves are mapped to the corresponding lie algebra space.All above is the feature extraction process of this paper.Thirdly,the method of human action recognition based on convolutional neural networks(CNNs)is studied.For the dimension of Lie group features extracted by using the Lie group skeletal representations is relatively high.For the purpose of better processing the high-dimensional data,reducing the complexity of recognition process and speeding up the calculation,this thesis adopted CNNs to learn and classify the action features.And we evaluate our method on Florence3D-action,MSR Action-Pairs and UT Kinect-Action dataset;the average recognition accuracy is 93.00%,93.68% and 97.96% respectively,which outforms some of the existing methods and achieves high recognition accuracy and robustness.Finally,in order to verify the better dataset transplantation ability and robustness of the method of this thesis,we build a human action dataset CQUPT-Action to test it,and the average recognition accuracy is 97.26%.In adition,the natural interaction application of virtual characters in virtual learning environment based on human action recognition is studied.Four testers have participated in the test of the virtual interactive system respectively,and the final test results show that the system is flexible and stable.