Perception And Analysis Of Human Motion Based On Wearable Multi Sensors

With the rapid development of Micro-Electro-Mechanical System(MEMS)in recent decades,motion sensing systems on human body based upon tiny wearable sensors have been significantly improved in miniaturization,light weight,low cost,high precision and high performance.These features promote the realization of wearable human motion perception and make it possible for the technique to walk into peoples' daily lives.Wearable sensorbased human motion perception can acquire time-domain and frequency-domain information during human motion,and then quantitatively and precisely analyze the details of human motion,which is objective,comprehensive,efficient,and accurate.It can provide accurate,diverse and intelligent services for more people,with a wide range of market demands and application prospects.This paper focuses on two major issues in the perception and analysis of wearable human motion,including the classification of human motion patterns,and the calculation of human trajectories and distances.The research involves the following points:Aiming at the problem of low classification accuracy of human motion state based on inertial sensors,a new feature extraction method is proposed to improve the human motion recognition algorithm in terms of time,precision and complexity.Firstly,an adaptive time window method is designed to divide the acceleration and angular velocity data acquired on the foot.The size of the time window is set according to the length of the period of each activity,rather than following a fixed length of time.The extracted features take full account of the periodic nature of the activity and can more accurately reflect the activity pattern,thereby improving the accuracy of the motion pattern recognition.Secondly,we propose a novel algorithm that combines the minimum redundancy,maximum correlation algorithm(m RMR)with feature weight algorithm(Relief F)to perform feature dimensionality reduction on the selected feature set.Finally,23 sets of optimal features were selected for support vector machine(SVM)training,and seven motion states were identified: static,walking,running,jumping,going upstairs,going downstairs,and falling.The final experimental results show that the algorithm in this article can achieve an average recognition rate of 96.4% in the above seven motion modes.Compared with other algorithms,the time consumption is reduced,the classification accuracy is improved,while the original features are preserved,and the interpretability is firm.To solve the problem of human motion trajectory and distance calculation,we primarily analyze the characteristics of motion on the foot,and combine the differences in the zero velocity point duration,the angular velocity with the acceleration variation of different types of activities to set different detection indexes,which achieves more accurate estimation on the zero velocity points in the step period.Then we set up different detection indicators to achieve more accurate estimation on the zero velocity points in the step period.Next weuse double integrals in combination with ZUPT to eliminate the accumulative errors caused by long-term integration operations and system drift,which in turn makes the distance calculation more accurate.Finally,we conducted an experiment to acquire data and use the algorithm proposed in this paper to measure it.With the comparison between the estimation of the proposed method and the ground truth,the accuracy and reliability of the trajectory and distance calculation algorithm proposed in this paper were verified.