Research On Activity Recognition Based On Smartphone Sensors

With the rapid development of mobile communication technology and microelectronic technology,smart phones are rapidly spreading.Today's smartphones are integrated with various types of sensors and have powerful environmental awareness capabilities.Compared with wearable sensors,smart phones are more portable,and data collection and preprocessing are more convenient.Therefore,human behavior recognition based on smartphone sensors has high research value.However,due to the complexity and diversity of human activities,many challenges remain.This thesis mainly studies human body actions based on smart phone sensors,Ten kinds of actions including four kinds of transition actions are identified with high accuracy,and basic actions are identified by deep learning method.Based on the non-periodic and indefinite length of the transition actions,this thesis proposes to use fuzzy approximate entropy to detect the starting point and the end point of the transition actions to realize the automatic detection and segmentation of the transition actions,the segmentation accuracy rate reaches 94.5%.This paper proposes a hierarchical behavior recognition model based on support vector machines and random forests,and proposes to select different features for different sub-classifiers in the model.In the behavior recognition of the self-collected data set containing 10 actions,an accuracy of 98.1% was achieved.Experiments show that the model has good cross-person recognition performance,and can effectively improve the recognition performance of imbalanced data sets.Based on bidirectional recurrent neural network and LSTM structure,a deep bidirectional LSTM behavior recognition network was constructed,and hyperparameters such as the number of network layers and the number of neurons are determined through experiments.Finally,94.1%,98.5% and 98.9% recognition accuracy were obtained on UCI-HAR,WISDM and self-collected data sets.It has reached a relatively advanced level of research at home and abroad,and proved the effectiveness of the bidirectional LSTM behavior recognition network constructed in this paper.