Research On Human Motion Recognition Technology Based On MEMS Inertial Sensor

With the development of The Times,human motion recognition has become a hot research direction,widely used in virtual reality,health care,sports training,game entertainment and professional motion analysis and other fields.In this paper,the human motion recognition technology based on MEMS inertial sensor is studied.Different from the traditional human motion recognition technology,it overcomes the disadvantages such as bulky capture structure,high environmental requirements,personal privacy leakage and large amount of computation,and has the characteristics of low cost,good practicability and small size.It makes people can well integrate the equipment into their daily life,and carry out the development of various applications such as human motion analysis and prediction,which has great value.In this subject,research is carried out from the following aspects:Firstly,according to the relevant theoretical knowledge of human anatomy,the axial plane,joint,joint Angle and joint motion form of the human body are analyzed.At the same time,the human body motion model is simplified and abstracted to establish the human bone motion model,and the human body motion is characterized by the posture Angle of the human body joint.By watching a large number of human movement videos and access to relevant information,the 40 basic movements of the human body are classified and defined.Secondly,the fixed positions of 11 MEMS inertial sensors in the human body were determined according to the establishment of the human bone motion model,and 14 representative human movements were selected and the acquisition requirements were specified.The data of human movements were collected by MEMS inertial sensors.The data is transmitted to PC through wireless module,and the emerging non-relational MongoDB database is used for data storage to build the human movement sample database,which is prepared for the follow-up research on human movement recognition.Again,the abrupt value and noise in the data are removed by means of preprocessing,and the data are unitized.According to the multi-dimensional attitude Angle data,10 time domain features were extracted by means of mean,variance and kurtosis.Meanwhile,wavelet packet method was used to extract the time-frequency domain features of 8 energy values reconstructed by variance.PC A method was used to reduce the features,which laid a foundation for subsequent classification and recognition.Finally,the support vector machine(SVM)optimized by particle swarm optimization(PSGA)was used to intelligently classify 14 kinds of actions,and the classification results and parameters were discussed under the time domain features,time-frequency domain features and fusion domain feature sets.The experimental results show that the recognition rates are 92.85%,90%and95.71%in the time domain,time frequency domain and fusion domain,respectively.The classification results of fusion domain feature set are better than those of time domain and time frequency domain feature set.The recognition rates of 14 kinds of movements in the fusion domain feature set are given.