Artificial Intelligence of Things (AIoT), which combines the advantages of both Artificial Intelligence and Internet of Things technologies, has become widely popular in recent years. In contrast to typical IoT setups, wherein devices collect and transfer data for processing at some other location, AIoT devices acquire data locally and in real-time, enabling them to make smart decisions. This technology has found extensive applications in intelligent manufacturing, smart home security, and healthcare monitoring.
In smart home AIoT technology, accurate human activity recognition is crucial. It helps smart devices identify various tasks, such as cooking and exercising. Based on this information, the AIoT system can tweak lighting or switch music automatically, thus improving user experience while also ensuring energy efficiency. In this context, WiFi-based motion recognition is quite promising: WiFi devices are ubiquitous, ensure privacy, and tend to be cost-effective.
Recently, in a novel research article, a team of researchers, led by Professor Gwanggil Jeon from the College of Information Technology at Incheon National University, South Korea, has come up with a new AIoT framework called multiple spectrogram fusion network (MSF-Net) for WiFi-based human activity recognition. Their findings were made available online on 13 May 2024 and published in Volume 11, Issue 24 of the IEEE Internet of Things Journalon 15 December 2024.
Prof. Jeon explains the motivation behind their research. "As a typical AIoT application, WiFi-based human activity recognition is becoming increasingly popular in smart homes. However, WiFi-based recognition often has unstable performance due to environmental interference. Our goal was to overcome this problem."
In this view, the researchers developed the robust deep learning framework MSF-Net, which achieves coarse as well as fine activity recognition via channel state information (CSI). MSF-Net has three main components: a dual-stream structure comprising short-time Fourier transform along with discrete wavelet transform, a transformer, and an attention-based fusion branch. While the dual-stream structure pinpoints abnormal information in CSI, the transformer extracts high-level features from the data efficiently. Lastly, the fusion branch boosts cross-model fusion.
The researchers performed experiments to validate the performance of their framework, finding that it achieves remarkable Cohen's Kappa scores of 91.82%, 69.76%, 85.91%, and 75.66% on SignFi, Widar3.0, UT-HAR, and NTU-HAR datasets, respectively. These values highlight the superior performance of MSF-Net compared to state-of-the-art techniques for WiFi data-based coarse and fine activity recognition.
"The multimodal frequency fusion technique has significantly improved accuracy and efficiency compared to existing technologies, increasing the possibility of practical applications. This research can be used in various fields such as smart homes, rehabilitation medicine, and care for the elderly. For instance, it can prevent falls by analyzing the user's movements and contribute to improving the quality of life by establishing a non-face-to-face health monitoring system," concludes Prof. Jeon.
Overall, activity recognition using WiFi, the convergence technology of IoT and AI proposed in this work, is expected to greatly improve people's lives through everyday convenience and safety!
