Mobile Phone Sensor Data Analysis And Research On Human Activity Recognition

Human activity recognition belongs to the field of pattern recognition and activity perception in the direction of computer research.It mainly obtains the signal data of specific activitys based on sensors,and then preprocesses these data to extract meaningful features.Finally,it uses the traditional machine learning model or deep learning model to classify or predict.Human activity recognition technology can be used in security monitoring,human-computer interaction,medical care and other fields.Known from previous studies,according to the different data source,related research can be roughly divided into human activity recognition based on video or image processing recognition,custom sensors,smart phone three categories,due to the rapid development of sensor technology and dependence in intelligent mobile phone in people's life and study on the convenience of using mobile phone is used to identify the human activity of the sensor research has become the current hot spot.This paper analyzes and studies human activity recognition based on the sensor data in smart phones.The main research contents of this paper are as follows:1.Build a human activity recognition systemOn the Android side of the sensor data acquisition and real-time of linear acceleration,and then the data stored in EXCLE file,send data to the server,the server processes the data file,on the server side using the first-order exponential smoothing the noise from the data file for Android end and complete classification model to evaluate the experiment.2.Construct the framework of HARC(Human Activity Recognition Chain)Under this framework,four supervised learning algorithms are trained and optimized,namely KNN(k-nearest Neighbor),RF(Random Forests),SVM(Support Vector Machine)and CRNN(Convolutional Recursive Neural Network).The performance of the algorithm was evaluated by 5 times cross validation,which was layered by time and then layered by subjects.The conclusions are as follows: in the time stratified cross validation classification experiment,the classification accuracy of the four algorithms is all above 94%,among which the CRNN algorithm has the highest classification accuracy,which can reach 99.3%.In the experiment of subject stratification and cross-validation,the classification performance of the unknown subjects was evaluated,and the accuracy of the four algorithms decreased,among which the accuracy of CRNN remained the highest,reaching 88.9%.The experimental results show that the CRNN model constructed in this study has a good recognition accuracy in human activity recognition and classification based on mobile phone sensor data.