Research On Human Activity Recognition Based On Smartphone Sensors

Human activity recognition is the process of identifying one or more activities from observations about human activities and environmental conditions.This process can learn advanced knowledge about human activities from the records,which can provide effective support for various health and medical applications.In recent years,sensor-based human activity recognition has developed rapidly.The approach of combining smartphone sensor data with machine learning techniques,due to its attractive application value,has become an important part of current research.However,due to the lack of solutions for different deployment platforms in the current study,model deployment remains difficult.Therefore,in this research,finding the appropriate solutions for model deployment was taken as the optimization goal,and the model selection and the development of the corresponding data processing pipelines were considered as the main work.Finally,three solutions based on machine learning models were proposed.(1)The process of human activity recognition based on traditional machine learning models was built.First,the median filter was used to remove the isolated noise in the raw data,and the data was denoised in the frequency domain according to the prior knowledge,and then the sliding window was used to segment human activity samples.Second,the features derived from the first-order difference of angular velocity in the frequency domain were extracted,and a feature set containing 643-dimension was built.Third,the model selection was assisted by the T-distributed stochastic neighbor embedding algorithm.Finally,the results showed that the support vector machine model with linear kernel function could take the classification result and the time consumption into account.This model obtained 96.143%,91.093%,and 93.727% for the F1 score on the three data sets,and its training time was 2.350 s,5.446 s,and 4.202 s respectively.Due to the large requirements of computing resources and time consumption,this solution is suitable for deployment platforms with rich computing resources.(2)In order to improve the training and prediction speed,feature selection was performed.First,the distribution difference of feature data in different categories was observed,and then 35-dimensional features including acceleration magnitude,the angle between axes and gravity were selected to form a feature subset.Second,traditional machine learning models were used to perform classification on the feature subset,the results showed that the support vector machine model with linear kernel function obtained 89.481%,82.907%,and 87.899% for the F1 score on the three data sets,and its training time was 0.288 s,0.581 s,and 0.483 s respectively.The time consumption of this solution was close to 1/10 of that on the original feature set,and it could achieve acceptable classification result.Therefore,this solution is applicable to the situation where the computing resources of the deployment platform are insufficient.(3)Neural network was applied to human activity recognition.First,recurrent neural network(RNN),convolutional neural network(CNN),and the model in which these two were fused were built.Second,these models were used to perform classification on the data sets,and the GPU was used for acceleration.The results showed that RNN models had the longest training and prediction time due to the difficulty of acceleration through the GPU.The longest training time belonged to the bidirectional long short-term memory(LSTM)model,which reached 3194.870 s.Therefore,RNN models were not suitable for model deployment.After the fusion of CNN and LSTM,the CNN-LSTM model obtained 94.285%,91.445%,and 92.658% for the F1 score on the three data sets,and its training time was 21.744 s,28.456 s,and 25.752 s respectively.Compared with the similar neural network models,the CNN-LSTM model could obtain the most stable recognition result,and its time consumption was close to 1/43 of the LSTM model due to the advantage of being accelerated by the GPU.Therefore,this solution is applicable to the situation where the computing platforms could be accelerated by GPU.This thesis compensated for the lack of deployment solutions for different computing platforms in human activity recognition.The development of data processing pipelines and the model selection were carried out to provide a theoretical framework for the deployment,and three corresponding solutions were proposed.These solutions have the potential for learning high-level knowledge of human activities for health and medical applications,such as human health record,disease prevention,and patient assistance.