Research On Human Activity Recognition Based On Accelerometer Sensor

As an important application scenario of Internet of Things(Io T),human activity recognition(HAR)has important potential value.For example,it can be used for daily care with special healthcare systems,as well as customized fitness guidance for obese people and the ability to enrich the smart home.With the development and advancement of sensor technology,the data acquisition problem that plagues human activity recognition has been effectively solved.In this context,this paper proposes an efficient human activity recognition method based on accelerometer sensor.The method designed in this paper mainly includes four parts: feature extraction,feature dimensionality reduction,activity model training and performance verification.In the feature extraction stage,the time-frequency analysis method S transform(ST)is used as the feature extraction method for the first time,and the average value of each row of the transformation matrix is taken to eliminate the influence of some abnormal points on the feature space.In the feature dimensionality reduction stage,a regularized supervised subspace learning method based on low rank idea is used,which makes the object function of the sample space coefficient matrix,the sample label information and the simulated noise interference to be optimized.The low-dimensional denoising subspace can be reconstructed from the sample data subject to noise interference.In the training stage of activity model,in order to verify the robustness of the above feature space to different classifiers,this paper uses support vector machine(SVM),nearest neighbor(NN)and Bagging as the recognition algorithm to train the activity model.In the performance verification phase,two different verification methods are employed and the performance of the activity model is evaluated using accuracy,recall,precision and F1 as evaluation indicators.Finally,the method proposed is evaluated on the public data sets of WISDM,SCUT and m Health.From the perspective of the above performance evaluation indicators,it is verified that the introduced approve has higher recognition performance than the existing algorithms.