Research On Human Activity Recognition Based On Three-axis Accelerometer

With the end of the development of sensors and the popularization of mobile devices,the sensors in mobile phones always record the activity data of the human body.The analysis of these data can discover the characteristics of human activity,which has important research significance.Human activity recognition research is currently based on machine learning.Traditional machine learning technology relies on manual feature extraction in the process of feature extraction,which has certain limitations and low accuracy.Deep learning methods have achieved good results in image,natural language processing and other fields.There are few studies on activity recognition.How to optimize the structure of human activity models to improve accuracy is a problem that needs to be studied.At this stage,there are many human activity recording software in mobile phones,but they cannot automatically identify and record activities according to the state of the human body.In response to the above problems,this article uses the three-axis accelerometer in the mobile phone to collect human activity data,combines the deep learning method to build and optimizes the human activity recognition model,and uses the mobile deep learning framework to transplant the model to the Android phone to realize the real-time human activity Recognition.The main research content of this paper reflects the following aspects:(1)Aiming at the problem that machine learning methods have limitations in feature extraction.This article constructs an improved CNN-LSTM model,uses the CNN model to extract signal features in human activity data,connects to the LSTM layer through the maximum pooling layer,and inputs it to the LSTM model to extract the timing features.The hyperparameters of the model are selected through multiple comparison experiments.Use the trained model to experiment on WISDM and the data set collected in this article and compare with traditional machine learning methods.Experiments show that the CNN-LSTM model constructed in this article can effectively identify features in human activity data,and the accuracy is higher than that of traditional machine learning methods.(2)Aiming at the problem of confusion of similar actions in the CNN-LSTM model,and further improving the recognition accuracy to simplify the model structure.Using multi-data fusion method,combined with ConvLSTM model to effectively extract the spatio-temporal information in the data.Using ConvLSTM model for each sensor data,the hyperparameters are optimized through multiple comparison experiments,and the obtained feature results are fused.Use the optimized model to perform experiments on UCI-HAR and the data set of this article,and obtain an accuracy of more than 95%,which can effectively recognize human movements(3)Aiming at the problem that applications in mobile phones cannot perform real-time activity recognition based on human body status.Use the Android application development platform to design and develop human activity recognition applications,use the mobile device deep learning framework to transplant the model into the application,and obtain data from the sensors in the Android phone to realize human activity recognition.After experimental testing,the human activity recognition application designed in this article can perform real-time activity recognition based on the current state of the human body.