Research On Human Activity Recognition Based On Smartphone

Human activity recognition(HAR)is an important subject in computer vision field.It is widely applied in health care,smart home,security monitoring,game controlling and other fields.Smart phone has developed rapidly in recent years.And it carries various of inertial sensors and enough computing power,which makes the HAR based on smart phones more and more concerned by researchers.In addition,the application of HAR based on smartphones often pays attention to the real-time performance and energy consumption while pursuing high accuracy.Therefore,this thesis mainly analyses and studies these two aspects.The details are as follows:Traditional RNN and LSTM networks can extract time-dependent features from samples,which is very suitable for HAR application based on smartphone sensors,but they have high time complexity.Aiming at this problem,a multi-layer parallel LSTM network is proposed.The model extracts features by sample spltting,parallel processing and feature fusion.It can not only extract time-dependent features,but also reduce the time complexity.Experiments show that the proposed network has higher accuracy than traditional machine learning algorithm,and has lower time complexity.It is very suitable for real-time smartphone applications.In the application of HAR based on smartphone,sampling rate of sensors is one of the key factors affecting the overall performance.The higher the sampling-rate,the higher the recognition accuracy,but the higher the energy consumption.Therefore,this thesis proposes an adaptive sampling rate algorithm based on reinforcement learning,which can reduce energy consumption as much as possible while guaranteeing the accuracy of model recognition.In the algorithm,the Activity Recognizer is responsible for feature extraction and activity classification from samples.The Sampling-rate Selector selects the next sample's sampling-rate according to the feature of the current sample.Finally,the Sampler completes the data acquisition.Experiments show that the proposed adaptive sampling-rate algorithm can flexibly switch the sampling rate in real time,and reduce energy consumption while ensuring the accuracy of recongition recognition.