| Summary: | This dissertation delves into the field of Human Activity Recognition(HAR), focusing on the intersection of deep learning and Wi-Fi technologies. Our main challenge is to implement HAR using Wi-Fi channel response, which is an evolving area where traditional HAR systems rely primarily on sensor-based data. By reviewing existing research, we noticed that research in this field is valuable, and the mature application of deep learning in the traditional HAR
field provides ideas for this project. This dissertation mainly explores the performance of neural network architectures such as LSTM, BiLSTM, CNN+GRU and Visual Transformers (ViT) in HAR tasks, and explores the feasibility of Wi-Fi as a non-traditional data source. The dissertation trained and tested several neural networks based on the Widar and NTU-Fi datasets, respectively, and optimized the problems that existed during the testing of the networks corresponding to different datasets, with the best model in this dissertation having an accuracy of 1.27 times higher than that of SenseFi. This research demonstrates the potential of combining Wi-Fi technology with deep learning to pave the way for smarter, more efficient systems in healthcare, home automation, and industrial environments, and marks a major advancement in human-computer interaction that will provide contributions to the field.
|
|---|