Skip to main content
Back

QUT Bakes Physics Into WiFi AI to Detect Falls Without Cameras

A new physics-guided TCN uses Doppler energy and subcarrier variance to match deep model accuracy at a fraction of the compute cost

Paper, Full Access
  • Physics-guided attention: A compact TCN uses Doppler energy and subcarrier variance as explicit attention signals, replacing the need for deeper architectures in WiFi activity recognition.
  • Two novel modules: Doppler-Energy Temporal Attention weights time steps by motion intensity; Variance-Driven Channel Attention weights WiFi subcarriers by how much they respond to movement.
  • Benchmarked on three datasets: UT-HAR (7 classes), NTU-Fi HAR (6 classes), and CSI-HAR (7 classes), outperforming deeper baselines on accuracy while using fewer parameters and FLOPs.
  • Core insight: Baking known physics into the model architecture can substitute for model depth, a principle transferable to other sensing domains.
  • Key use cases: Fall detection, smart home automation, rehabilitation monitoring, and occupancy sensing -- all using existing WiFi hardware with no cameras or wearables.
  • No code release: The paper is on arXiv but no public code or pretrained model has been released alongside it.

Most WiFi-based activity recognition systems follow a familiar playbook: stack more layers, add more parameters, and let the model figure out what matters in the signal. A new paper from researchers at Queensland University of Technology argues that this brute-force approach is unnecessary -- and that baking physics directly into the model's attention mechanisms can achieve better results with far fewer resources.

The invisible sensor hiding in your router

Human activity recognition based on Channel State Information (CSI) using commercial WiFi devices plays an increasingly important role in many applications, such as smart home and interactive games. The core idea is elegant: when a person moves through a room, their body disturbs the WiFi signal traveling between a router and a receiver. CSI describes the channel properties of a communication link, reflecting how wireless signals propagate in a physical environment when subject to obstacles, reflection, diffraction, and scattering. For each pair of antennas, CSI captures the wireless channel characteristics, including time delay, amplitude attenuation, and phase shift of multipath components on each communication subcarrier. By reading those disturbances, a model can infer what the person was doing -- walking, falling, sitting down -- without any camera or wearable device.

Vision-based and sensor-based methods can present better data but at the cost of users' inconvenience and social constraints such as privacy issues. WiFi sensing sidesteps all of that. This method is advantageous for its non-intrusiveness, cost-effectiveness, and ability to work through walls and in various lighting conditions. The catch is that CSI signals are noisy, environment-sensitive, and packed with subcarriers that may or may not carry useful information about the activity at hand.

The problem with going deeper

Existing deep learning-based HAR approaches usually extract either temporal or spatial features and lack adequate means to integrate and utilize the two simultaneously, making it challenging to recognize different activities accurately. The field's response has generally been to add complexity: deeper CNNs, stacked LSTMs, transformer-based architectures. Previous works have relied on heavy and deep models to achieve performance improvements. Although effective, these models entail increased computational complexity and longer inference times.