Hnoohom N.Mekruksavanich S.Theeramunkong T.Jitpattanakul A.Mahidol University2024-09-202024-09-202024-04-12ACM International Conference Proceeding Series (2024) , 113-119https://repository.li.mahidol.ac.th/handle/20.500.14594/101277The use of deep learning (DL) technology for the purpose of human activity recognition (HAR) is an important research area. Vision and sensor-based methods can provide good data but at the cost of privacy and convenience. Furthermore, Wi-Fi-based sensing has become popular for collecting human activity data, as it is ubiquitous, versatile, and performs well. The utilization of channel state information (CSI) obtained from Wi-Fi networks has the potential to facilitate the recognized activities. Traditional machine learning relies on hand-crafted features, but DL is more appropriate for automated feature extraction from raw CSI data. This work presented a generic HAR framework using CSI and studied various deep networks. We proposed a deep residual network that would automatically extract informative features from raw CSI. In this study, we conducted a comparative analysis of five fundamental deep networks; namely, convolutional neural network (CNN), long short-Term memory (LSTM), bidirectional LSTM, gated recurrent unit (GRU), and bidirectional GRU. Experiments on a publicly benchmark dataset named the CSI-HAR dataset showed that the proposed recognition model performed the best for CSI-based HAR with the highest accuracy of 98.60%, thus improving the accuracy by up to 3.60% over prior methods. Therefore, deep residual networks would be considered to be a suitable option for HAR tasks that would encompass Wi-Fi CSI data.Computer ScienceConference PaperSCOPUS10.1145/3664934.36649502-s2.0-85203871067