Research On The Wearable Sensors-based Human Activity Recognition

With the continuous development of microelectromechanical systems,the human behavior recognition system based on wearable sensors has very important applications in various fields such as health management,medical assistance,sports analysis,military and industry.In view of the problem of large noise in accelerometer data and cumulative error of gyroscope sensors,this dissertation takes the azimuth information of inertial navigation system as an entry point,and studies the practicability of azimuth information to improve the accuracy of human behavior recognition system.The main aspects of this study include:In view of the large noise of the accelerometer,the azimuth is proposed to be applied to human behavior recognition system as a new feature group.At present,most studies only use one or a very small part of the features extracted from azimuths.The vast majority of researches still focus on human behavior classification based on triaxial acceleration data.In this paper,azimuth information is taken as a new feature group earlier,and the system is comprehensively used to comprehensively evaluate the impact of this feature group on the performance of human behavior recognition system.Aiming at the problem of long-term measurement of gyroscope sensor's cumulative error,an azimuth solving algorithm based on complementary filtering and Kalman filter is proposed.An azimuth solving algorithm is proposed to correct the azimuth using the gyroscope for azimuth,accelerometer and magnetometer.It can be seen from the comparison of the results that the algorithm successfully improves the azimuth accuracy and provides a guarantee for obtaining the accurate azimuth characteristics subsequently.In order to solve the problem of low computational complexity of recognizing human behavior recognition system,after obtaining the accurate azimuth information,this paper extracts the features of the time-domain,frequency-domain and time-frequency features of different behaviors and applies the maximum correlation minimum redundancy algorithm Filter features.Finally,we use typical classification models: Support Vector Machine(SVM),K-Nearest Neighbor Algorithm(KNN)and Naive Bayes Algorithm(NBC)to classify features to reduce computational complexity and ensure recognition efficiency.