Human Activity Recognition Based On Machine Learning

The technology of Human Activities Recognition(HAR)based on sensors has a wide application prospect in the fields of human-computer interaction and medical care.How to select and improve HAR algorithm is the research hotspot in HAR field.At present,the classification algorithms used in the field of HAR based on acceleration sensors mainly include traditional machine learning algorithms and deep learning algorithms.The recognition process of these two algorithms is consistent.Firstly,the acceleration data are preprocessed,and then the labeled data are used to train the recognition algorithms.Finally,the acceleration data are classified by the training results of the algorithm.For now,the collection standards of many public datasets are inconsistent,so it is necessary to collect data and build datasets by ourselves in practical application.In addition,the current HAR researches mostly stay in the algorithm verification stage,and there are few HAR studies which deploy their training results.In this paper,the optimal algorithm training results are deployed into Android phones,which can accurately recognize human activities in real-time.The main research contents of this paper are as follows:(1)The triaxial acceleration data generated by human activities were collected and processed.By using the built-in triaxial acceleration sensors of Android phones,X,Y and Z triaxial acceleration data of seven kinds of human activities were collected at the sampling frequency of 50 Hz,and the data were grouped and filtered,then their features were calculated to establish a measurement data set for algorithm training.The acceleration data of Android phones worn at waist and legs were collected to construct two datasets.(2)Two HAR algorithms,multi-layer recognition and LSTM-1DCNN were proposed.A multi-layer recognition algorithm was proposed based on the thinking of decision tree.The LSTM-1DCNN recognition algorithm was proposed,which combines LSTM and 1DCNN in parallel to extract features from different angles.Using five public datasets and two self-built datasets as the original data,we compared the performance indicators of decision tree,random forest,SVM,multi-layer recognition,1DCNN,LSTM and LSTM-1DCNN on different datasets.After comprehensive comparison of the performance indicators of different algorithms,LSTM-1DCNN were finally selected as the deployment algorithm.(3)The overall design of mobile Android applications was completed and validated.The work includes data collection,processing,algorithm forward calculation and result conversion on Android mobile phones,and the actual testing and result analysis was conducted.Actual testing shows that the LSTM-1DCNN recognition algorithm constructed in this paper achieves an overall recognition rate of 97.76% for seven human activities when running on Android phones.