WiCNNAct: Wi-Fi-Based Human Activity Recognition Utilizing Deep Learning on the Edge Computing Devices

Nowadays, human activity recognition plays an essential role in the application of human-computer interaction. Comprehensive systems, however, mostly rely on wearables, video cameras, and ambient sensors, which might be expensive and difficult to deploy or cause privacy issues. Wi-Fi-based wireless human sensing utilizes channel state information to detect human presence and movements, leveraging changes in signal properties caused by humans. This cost-effective approach utilizes existing Wi-Fi infrastructure and can operate in obscured or obstructed environments. However, signal interference and environmental conditions can affect accuracy and reliability. To address these challenges, a WiCNNAct for human activity recognition is proposed in this paper. The proposed approach utilizes the channel state information (CSI) measurements (complex values) from Wi-Fi and processes the different combinations of the real, imaginary, and absolute values using multi-channel 1D convolutional neural networks (1D-CNN). After conducting preliminary investigations, we validated various combinations of multi-channel 1D-CNNs and identified three methods for accurate activity recognition: a three-channel (real, imaginary, and absolute) setup, two channels (real and imaginary/real and absolute/imaginary and absolute), single channel (real/imaginary/absolute). The proposed three-channel Wi-Fi-Net underwent verification with 10-fold testing validation and achieved an overall accuracy of 98.29% with a standard deviation of 0.33%. The model is deployed on various edge computing devices, including Raspberry Pi, to assess real-time deployment feasibility. In this work, 1D-CNN is chosen based on its ability to automatically extract spatial and temporal features, reducing manual feature engineering. Experimental validation identifies the optimal setup for robust activity recognition.