Research Of Daily Activity Recognition Methods Based On Sensor Data And Deep Learning

With the development of the Internet of Things and pervasive computing,human activity recognition has become one of the most important issues in pervasive computing.At the same time,because of personal privacy,researchers are popular with the human daily activities recognition based on sensor data increasingly.In recent years,deep learning and the rapid development of hardware device GPU acceleration have made great breakthroughs in many computer application contexts.Therefore,this paper focuses on the application of sensor data and deep learning methods in the field of activity recognition.However,it's difficult to recognize activities of user Bs with the model trained for user A.The effect of transferring the model(among different users)is the key that restricts activity recognition in practice.At present,there is still little research on the transferring of deep learning model in this field.Its effect,principle and influencial factors remain to be studied.Therefore,we carried out the empirical study on the transferring of deep learning model among users.We visualized the features extracted from CNN and studied its distribution.We compared the feasibility,strengths and weaknesses of typical unsupervised and semi-supervised transferring methods.Finally,this paper proposed an unsupervised transferring method in the field of activity recognition.In the research part of sensor-based activity recognition method,this paper can be divided into three parts: activity recognition method based on discrimitive model,activity recognition method based on generative model,and activity identification method based on hybrid model.This paper carried out experimental verification on three public datasets,and proposed a CNN-LSTM deep learning framework that is common in the field of activity recognition,and paved the way for subsequent studies on transfer learning methods.In the part of deep transferring experimental research in the field of activity recognition,this paper adopted the unsupervised transferring method of DANN and the semi-supervised transferring method of Fine-tune.Through CNN visualization source domain training model to recognize the target domain directly,this paper found that the features extracted by CNN have the small distance between classes and large discrepancy within the same class,which made it difficult to recognize.This paper hopes to further narrow distance within the same class based on the transferring method among different people.At the same time,to expand the distance between classes for improving the accuracy of the target domain.Therefore,this paper proposed a method based on Centorloss and MMD loss function for unsupervised transfer between different people in the field of activity recognition.At the same time,this paper adopted the common unsupervised transferring method to make comparative experiments on three public data sets.In addition,in the real-world environment,three-axis acceleration data of five different activities(upstairs,downstairs,sitting,standing,walking)were collected using an acceleration sensor of an Android mobile phone.Then the three-axis acceleration data in the field of activity recognition was visualized.At the same time,a combination of convolutional neural network and transferring method was used.Experiments were carried out for three individuals with active transfer(six cases).Through experiments,the transferring method of Center Loss and MMD loss function presented in this paper can effectively improve the accuracy of the target domain.In the field of activity recognition,it has certain advantages over other unsupervised transferring methods.Finally,this paper visualized the feature distribution and confusion matrix of the target domain,and further verified the transferring method of Center Loss and MMD loss function.