Research On Human Activity Recognition Algorithms For Smartphones

Human activity recognition is the process of sensing human activity data through various sensors,and using computer automatic detection technology to analyze and understand various activities of the human body.It has broad application prospects in many aspects,including video surveillance,medical diagnosis and monitoring,smart furniture and intelligent human-computer interaction.In the aspect of human activity recognition,the current research work mainly uses traditional deep learning models for feature extraction,which has problems of insufficient extraction of key features,large workload and difficulty in ensuring timeliness.Therefore,the research of efficient and accurate human activity recognition algorithm has become a new hotspot.This paper aims to explore different activity recognition algorithms,and select the algorithm with high efficiency,real-time recognition and generalization ability for human activity recognition.The core of human activity recognition algorithms mainly includes perception and recognition.It aims to effectively integrate advanced pattern recognition technology,sensor technology and computer technology into the whole human activity recognition system.This paper comprehensively analyzes and summarizes different deep learning algorithms,explores the performance of different deep learning algorithms in feature extraction on time series data.We design two novel models to extract fine-grained features of human activity recognition,which greatly improve the performance of activity classification.In addition,this paper also comprehensively analyzes the characteristics of sensor data and explores its application value in human activity recognition system.This paper has made a number of innovations and achievements,the main contributions include:(1)We propose a novel human activity recognition network HDL,which combines DBLSTM(Deep Bidirectional Long Short-Term Memory)model and CNN(Convolutional neural network)model.In order to better capture the information of time series data,we use the DBLSTM model which consists of many BLSTM layer.The output of current BLSTM layer is transmitted to the next BLSTM layer.Each BLSTM layer which has an feature fusion can obtain more context information.As the BLSTM layer increases,the bidirectional output can be more deeply integrated.Therefore,we can get fine-grained feature representations at the last level.In addition,we introduce CNN into the DBLSTM model to extract features,which achieves remarkable results.The experimental results show that the proposed HDL network achieves reliable results with accuracy and F1 score as high as 97.95% and 97.27%.Compared with other networks based on the same smartphone dataset,the accuracy of HDL is higher than S-LSTM and Dropout CNN network by 2.14% and 6.97%respectively.(2)HDL for feature extraction just simply filters the data in order,without capturing key features of response activity changes.Therefore,we design a novel human activity recognition model ABLSTM which combines BLSTM and attention mechanism in this paper.BLSTM which connects the forward LSTM and the backward LSTM is used to extract coarse-grained features on the input data.Attention mechanism is used to analyze the important degree of the coarse-grained features set to obtain fine-grained features which are more salient for activity recognition.ABLSTM can be used not only for processing of the time series data,but also for"attentive features " which are in response to the changes of human activities.By testing our model on two public benchmark datasets,UCI dataset and Opportunity dataset,the results show that our model can well identify human activities with F1 score as high as 99.0% and 92.7%respectively,which pushes the state-of-the-art in human activity recognition of mobile sensing.(3)We verify our models on the real smartphone dataset.Different types of sensors are built into smartphones,and the data collected by sensors have the characteristics of time sequential.The distributed time series data are used as the input of human activity recognition model,and the context information of time series data is used to find out when time series data is relevant,instead of simply filtering data.This can effectively extract the data features that changes with time,and make up for the loss caused by the feature extraction of image data in the past,which improves the accuracy of activity recognition.