Research On Cross-domain Human Activity Recognition Method Based On Transfer Learning

Human Activity Recognition(HAR)is an important research content in the field of pervasive computing.It has been widely used and developed in children’s movement monitoring,patient rehabilitation training,and elderly fall detection.At present,the research on behavior recognition methods based on deep learning has attracted the attention of many scholars,but most deep learning models are designed to solve specific tasks.When the data distribution changes,these models will require a lot of computing power and spend a lot of time to be reconstruction.And transfer learning can use a pre-trained network and apply it to our custom tasks,as well as transfer the knowledge learned from previous tasks.Therefore,this thesis studies the cross-domain recognition method of human behavior from two aspects:parameter migration and feature migration.The main contents include:(1)Human activity recognition generally includes the steps of action data collection,preprocessing,model training and recognition.In this thesis,aiming at the problem of non-specific action category migration and recognition,target domain data without labels,etc..Based on Convolutional Neural Networks(CNN)and Long Short-Term Memory(LSTM),a migration-based learning is designed.Framework of cross-domain behavior identification methods.At the same time,for the problem of motion data segmentation,a motion data segmentation method based on change point detection is proposed,which can achieve accurate segmentation of different motion data in continuous motion data and provide input data for model training.(2)Proposed an activity recognition method based on parameter migration.In view of the changes in the overall data distribution caused by the appearance of new behaviors in the target domain,which causes the problem of model performance degradation.Using the source domain samples to train the HAR model based on CNN-LSTM firstly,and apply its cross-domain migration to the target domain.Then,when the target domain sample changes or a new behavior category appears,the parameters of the CNN and LSTM feature extraction layers in the HAR model are frozen,the network parameters are retained,and the fully connected classification layer of the model is retrained.Obtained HAR model updated through parameter migration learning.The new HAR model maintains the ability to recognize specific actions in the source domain,while increasing the types of action recognition,and it can identify new behaviors in the target domain at the same time.Experimental results show that the method can gradually add the ability to recognize new behaviors in the target domain while maintaining the model’s existing recognition accuracy of specific actions in the source domain,and has good recognition accuracy.(3)Proposed an activity recognition method based on feature transfer.Aiming at the problem of missing labels in the sample data collected in the target domain.Firstly,the unsupervised source domain selection method is used to comprehensively measure the relevant distance between the target domain and different source domains,and select the source domain with the highest correlation for the target domain.Secondly,using DAN(Deep Adaptation Network)and Multi-Kernel Maximum Mean Discrepancy(MK-MMD)methods to minimize the overall distribution difference of the two domains.Cross-domain transfer of the model trained in the source domain to the target domain.The experimental results show that when there is a large amount of unlabeled data in the target domain,the migrated model can be used to label the target domain without labels.In addition,when the source domain data and the target domain data obey different distributions,this method can align the overall distribution difference between the source domain and the target domain,increase the adaptability of the model to the target domain samples after the cross-domain migration,and it will not cause a significant decrease in the performance of the model,and also avoid a lot of time spent retraining a new model.