Research Of Human Motion Pattern Recognition Method Based On MIMU

Human motion pattern recognition based on Micro-Inertial Measurement Unit(MIMU)is a new research direction in the field of pattern recognition.MIMU has the advantages of small size,light weight and wearability,and can be uesd to implement the continuous monitoring of motion pattern.Therefore,MIMU has broad application prospects in the fields of medical health,sports and military training.Recognition of eight kinds of daily motion patterns are researched in this paper,including standing,sitting,lying on back,walking,ascedning,descending,running and cycling.The main work of this dissertation are as follows:(1)Aiming at high performance,low power consumption,miniaturization and low cost,the device selection and hardware circuit design of sensor nodes are carried out;the motion characteristics of each motion mode are analyzed by human motion model,and the body parts with the best mode discrimination ability are determined,and the deployment scheme of sensor nodes is designed.(2)Aiming at the problem of low-cost MIMU with low accuracy and no calibration conditions for high-precision instruments,a mathematical model of MIMU system error is established.The error components such as coordinate system misalignment,coupling between axes,scale factor and zero offset are taken into account in the model.A software calibration algorithm based on multi-position method is designed according to the error model.The MIMU error calibration experiment is carried out according to the algorithm flow.The experimental results show that the output error of accelerometer after calibration is 74.1% lower than that before calibration,and the output error of gyroscope is 32.0% lower than that before calibration.(3)Aiming at the problem of low quality of raw data,combined with the knowledge of data mining field and the background of this research,outlier correction method based on orthogonal distribution and data missing processing method based on mean interpolation are designed.Ensembel Empirical Mode Decomposition(EEMD)is proposed to reduce the noise of sensor signals,which is prior to dependence.EEMD has self-adaptive characteristics based on wavelet transform of check basis function.Motion pattern classification experiments are carried out on the original data set and the data set processed by the data preprocessing method designed in this paper.The experimental results show that the average recognition rate of the hierarchical algorithm is improved by 2.4% after data preprocessing,and the average recognition rate of LSTM classifier is increased by 2.3%.(4)A hierarchical motion pattern recognition algorithm based on Support Vector Machine(SVM)is designed,and the feature scheme is designed according to the ability to distinguish the data set;the correlation features are extracted according to the motion characteristics of cycling;and the EEMD features are extracted for three confusing patterns of walking,going upstairs and going downstairs.Experiments show that the average recognition rate of the hierarchical algorithm for eight motion patterns is 96.7% in this data set.(5)A motion pattern classifier based on Long Short-Term Memory(LSTM)is designed.Compared with traditional machine learning which requires manual feature engineering based on expert knowledge,LSTM can automatically extract and fuse features of input time series,and further learn through hidden layer network.Experiments show that the average recognition rate of LSTM classifier for eight motion patterns is 97.6% in this data set.(6)In order to further validate the proposed algorithms,recognition experiments are carried out on open source DSAD datasets of UC Irvine.The experimental results show that the hierarchical algorithm based on SVM has good generalization ability,while the classifier based on LSTM has higher performance upper limit,but needs large-scale training datasets as support.