摘要 : Wi-Fi based human activity recognition is emerging and preferable over other approaches due to its numerous advantages, including privacy considerations, ubiquitousness, and easy deployment. While existing literature predominantly... 展开 Wi-Fi based human activity recognition is emerging and preferable over other approaches due to its numerous advantages, including privacy considerations, ubiquitousness, and easy deployment. While existing literature predominantly focuses on identifying the activities of a single-user, recognizing multi-person interactions (MPIs) is increasingly significant due to their profound social implications. However, research in this area has not progressed due to the limitation of publicly available Wi-Fi datasets and the complexities of MPI recognition. Motivated by this, we develop and publicly release a Wi-Fi channel state information (CSI) based MPI recognition dataset, coined Wi-MIR, that uses three transmit and three receive antennas to capture 270 ( $ {3}\times {3}\times {30}$ ) subcarriers with a sampling rate of 950 Hz. This dataset consists of 3740 trials encompassing seventeen distinct MPIs and is collected by eleven human pairs in an indoor environment. We put forth a lightweight deep learning model with attention mechanisms for MPI recognition from CSI, named CSI-IRNet, that adeptly concentrates on pertinent features, filtering out irrelevant elements, and mitigating the impact of signal complexity within the CSI for recognizing MPIs accurately. In addition, we compare the developed Wi-MIR and the existing public dataset by evaluating the performance of MPI recognition on both datasets to highlight the strengths and advancements provided by Wi-MIR. The evaluation results show that Wi-MIR dataset demonstrates a superior recognition performance by utilizing more subcarriers with a higher sampling rate as well as covering more diverse kinds of MPIs (Bowing, Conversation, Exchanging objects, Helping standup, Helping walk, and Touching another person). 收起