Research And Implementation Of Low-cost Domain Adaption Towards Cross-scene Wireless Sensing

With the development and popularization of wireless networks,researches and applications for wireless sensing are increasing.Compare with other sensing schemes,WiFi-based solutions have the advantages of low-cost,no portability,and privacy protection.However,due to instability and sensitivity to the environment of WiFi,the machine learning based on supervised learning suffers from significant performance degradation when the environment changes.To solve this problem,most of the existing work requires part of labeled data or abundant unlabeled data in the target domain to realize domain alignment or data generation,which has a high cost of domain adaptation.Therefore,this thesis studies the impact of environment changes on WiFi sensing,proposes a low-cost domain adaptation method,and conducts cross-scene verification through gesture recognition and activity recognition.The main work of this thesis is listed as follows:（1）This thesis proposes a low-cost domain adaption method based on virtual data generation for environment changes,utilizing source domain samples and one sample per action in the target domain to synthesize reliable virtual samples for the target domain.Firstly,the source data and target data belonging to the same class are given adaptive domain weights to achieve liner transformation,and then enhance the target domain information for the new sample from the linear transformation by the generate coder to obtain virtual samples in target domain,in which the adaptive domain weights are automatically determined by model training.In this thesis,the dataset of gesture recognition and activity recognition dataset are used for verification,conduct cross-scene experiments on user change,location change and room change,and achieve better results higher from 6.48% to 15.13% than the state of art.（2）The feature-level fusion is applied to amplitude and phase to improve the accuracy for cross-scene.Propose a feature-level fusion method based on consistency constraint and action semantic constraint,and extract action feature from amplitude and phase,then train the fusion model and classifier alternately.Using source domain samples and one sample per action in the target domain to conduct experiments,for cross-subject gesture recognition,the results in two rooms are 91.67% and 85.53% for target subject;for cross-location activity recognition,the results in two rooms are 90.53% and 85.03%for target location;using dataset from different environments for cross-room,the results for gesture recognition are 88.01% and 83.04% for target room,the results for activity recognition are 87.9% and 82.4% for target room.（3）Design and implement a WiFi-based cross-scene action recognition system.Based on the low-cost domain adaptation method proposed in this thesis,the action recognition model can be established quickly for a new scene.The data collect,adaptation model training,real-time monitoring,et al.are designed,implemented and tested.The system can realize low-cost domain adaptation and real-time motion monitoring for a new scene.This thesis proposes a low-cost domain adaptation method based on data generation and a fusion method for amplitude and phase to increase the accuracy for domain adaption.The method is common for scene changes in gesture recognition and activity recognition,and also suitable for room adaption with multiple scene changes simultaneously.At last,design and implement a WiFi-based cross-scene action recognition system based on the proposed method.