Research On Human Activity Recognition Based On Smartphone Sensors Data

The increasingly rich sensor configuration of smart phones makes the recognition of human activities on mobile platforms more and more common and direct.With the improvement of living standards in recent years,the demand for activity recognition systems applied to the fields of medical care,entertainment,and assisted living has greatly increased.Although research on human activity recognition has made progress,there are still many challenges due to the complexity of human activities.The success of deep learning methods in other machine learning areas has made it more and more concerned in the field of activity recognition.This paper focuses on the research of smart phone sensor-based activity recognition.The goal is to use a deep learning method to train a robust network model,and then transplant the model to the Android platform for real-time activity recognition.First,this paper studies the activity recognition based on convolutional neural network.The layer structure and training process of convolutional neural network are introduced,and the network model of convolutional neural network plus statistical features is constructed.Through several comparison experiments,the influence of the main hyperparameters such as the number of convolution layers and the number of convolution kernels on the model was analyzed.Experiments were performed on UCI-HAR,WISDM,and the dataset collected in this paper,the experimental results were compared with other traditional methods.Experiments show that the convolutional neural network model constructed in this paper can effectively automatically learn sensor data features and has a good recognition effect.Secondly,the research on activity recognition based on the recurrent neural network is carried out.The structure and training process of the recurrent neural network are introduced,and the network model is constructed using its deformed LSTM network.The main parameters such as LSTM layer and sliding window length were compared and tested several times,and the best parameter values were selected for the model.Experiments show that the LSTM network model can effectively learn the long-term dependence of sensor data,and the activity recognition effect is better than the traditional method.Finally,using Android Studio to design an Android real-time activity recognition application,the trained convolutional neural network model or LSTM network model is transplanted to the application for activity identification.Through testing,the activity identification application designed in this paper can better identify the activities that humans are performing.