Human Activity Recognition Based On Wi-Fi Noisy Label Data

Human activity recognition technology aims to make full use of advanced communication technology,artificial intelligence algorithms and other technologies to achieve real-time,and accurate human activity recognition systems.It is an important component of human-machine interaction and has important applications in medical monitoring,smart homes,and motion sensing games.Combining the widely deployed Wi-Fi devices and rapidly developing deep learning technology,Wi-Fi-based human activity recognition technology has attracted attention due to its low cost,noninvasiveness,and wide coverage compared to traditional activity recognition methods.However,the success of deep learning depends on large-scale datasets,and noisy labels caused by labeling errors in practice are inevitable,which will have a negative impact on model performance.Therefore,how to robustly achieve human activity recognition on Wi-Fi noisy label dataset is an important issue.To improve the recognition accuracy and robustness of deep learning models in WiFi noisy label data and ensure better practical performance of the model,this thesis conducts the following research:1.The UT-HAR and NTU-HAR datasets,which map human body behavior using Channel State Information(CSI),were used to create raw CSI datasets.Meanwhile,different levels of noisy CSI datasets were constructed using noise transfer matrices.Popular models for human activity recognition based on Wi-Fi data were introduced and modeled.Experiments were conducted on both the raw CSI dataset and the noisy CSI dataset,which clearly showed that the noise labels in the dataset had a negative impact on the models performance.2.This thesis handles noisy labels from the perspective of small loss instance selection.First,two models with the same structure but different initializations are maintained through Co-teaching.The two equivalent networks select small loss instances and update network parameters mutually.However,this training method requires the model to be reliable enough when selecting small loss instances,otherwise the model performance will further deteriorate.To address this problem,this thesis proposes a joint loss function based on contrastive learning and co-regularization,which maximizes the consistency between the two models and increases confidence when selecting small loss instances.Meanwhile,two training methods are obtained according to the function’s working stage.Through extensive experiments,it can be found that these two training methods can effectively combat noisy labels and improve the recognition accuracy and robustness of the model on the noisy CSI dataset.