Research And Implementation Of Human Activity Recognition And Its Transfer Learning Model

In this thesis,human activity recognition(HAR)technology refers to the process of signal processing,analysis and modeling of the quasi-periodic time series collected by sensors when the subjects do some activities,which can eventually recognize different actions.It is the basic supporting technology in many fields,such as health,medical treatment,assistant robot research.Especially,the activity recognition system based on portable devices has a wider application.In the field of HAR,adaptability is an urgent problem to be solved,that is,how to guarantee the recognition performance of the model in dynamic environment.The acceleration data acquired by smart phones during human motion are used for data processing and modeling.The emphasis is to improve the generalization and adaptability of the recognition model in real dynamic environment.The specific work is as follows:(1)In the part of data processing and feature extraction,to solve the problem of inconsistent length of motion units and unbalanced samples,based on sliding time window method,the corresponding data expansion scheme is proposed.in addition,aiming at the randomness of the placement direction of smart devices,a directionindependent feature vector is proposed by extracting the relative angle between the instantaneous acceleration and the average acceleration,which eliminating the uncertainty of the placement direction of sensors.(2)In the part of modeling,aiming at the problem that the traditional manual extraction features cannot recognize the upstairs,downstairs and walking well,a deep convolution HAR-Net model suitable for the field of HAR is established.An end-to-end recognition model with user independence is realized,and the model is further explained by visualization analysis.(3)In the part of model transferring,to meet the requirement of high adaptability and scalability of activity recognition model,a deep transfer HAR-Net model based on Fine-tuning is established to solve three supervised transfer learning tasks in the target domain.In addition,a deep transfer HAR-Net model based on DDC network is established to solve unsupervised transfer learning task in target domain.On this basis,a two-stage HAR-Net transfer model is proposed,which further improves the recognition performance of the unsupervised task.(4)In the part of system implementation,an app for acceleration data acquisition and real-time activity recognition based on Android is realized,and the performance of the model in real environment has been tested.The results show that the average recognition accuracy of the deep convolution HAR-Net model is 98.94% in the experimental environment and 97.97% in the real environment.In addition,deep transfer HAR-Net model can complete model transfer in different application scenarios with a short training time under the condition of small samples,and the transfer effect is good.When the amount of data in the target domain is only 20% of the source domain,the enhancement effect of supervised transfer tasks in the target domain is averagely more than 35%,and that of unsupervised transfer task in the target domain is about 34%.Finally,this paper summarizes the selection principles of transfer learning methods in the field of sensor-based HAR for different transfer purposes.