This paper presents Wi-HAR, a framework that turns commodity Wi-Fi devices into sophisticated activity sensors. By leveraging multi-antenna Variational Auto-Encoders (VAEs) and evidential deep learning, the authors achieve high-accuracy activity recognition while solving a critical real-world problem: detecting when the sensor's environment has changed enough to make its predictions unreliable.
TL;DR
Wi-HAR uses a multi-antenna VAE architecture to achieve 96.6% activity recognition accuracy using standard Wi-Fi signals. Crucially, it employs evidential uncertainty to detect "deployment shifts"—like moving the receiver or changing the room—flagging 82.5% of environmental changes without ever being trained on shifted data.
What Problem Does This Paper Address?
Wi-Fi sensing relies on Channel State Information (CSI), which captures how human movement perturbs wireless signals. While powerful, CSI is notoriously "brittle." A model trained to recognize walking in a specific lab might fail if the receiver is moved six inches, if a different person enters the room, or if furniture is rearranged.
Most existing models fail silently: they continue to output high-confidence activity labels even when the underlying physical environment no longer matches their training data. This paper tackles the dual challenge of achieving high recognition accuracy while providing a robust "check engine light"—an uncertainty signal that alerts users when deployment conditions have shifted.
Methods
The authors propose a multi-antenna architecture called Wi-HAR. Its key components include:
- Antenna-Specific VAEs: Instead of one large model, they use four parallel VAEs—one for each receiving antenna. Each VAE compresses raw Wi-Fi 6 CSI magnitudes into a very compact "latent" representation (only 2 to 12 parameters).
- Delayed Fusion: The means and log-variances from these four latent spaces are concatenated and fed into a Multi-Layer Perceptron (MLP) classifier. This allows the model to exploit the fact that physically separated antennas see "complementary views" of the same activity.
- Evidential Deep Learning (EDL): To handle uncertainty, the authors replace the standard Softmax output with an evidential head. This treats the network's output as parameters of a Dirichlet distribution, allowing the model to quantify its "lack of evidence" for a prediction.
- Class-Conditional Anomaly Scoring: They don't just look at raw uncertainty. They compare a new sample's uncertainty to the distribution of uncertainty seen for that specific activity during training. If the uncertainty is "unusual" for a "Walking" label, the system flags a deployment shift.
Datasets
The study utilizes a public Wi-Fi 6 CSI dataset recorded in a 46-square-meter indoor environment (see Figure 1).
- Baseline (S1a): Participant A performing 12 activities (Walk, Run, Sit, etc.) in a controlled lab setting.
- Shifted Datasets: Eleven distinct scenarios designed to break the model, including:
- Changing the receiver position.
- Replacing Participant A with Participant B.
- Recording in a different room (an office).
- Recording at a different time of day.
Key Insights and Ideas
- The Power of Multi-Antenna Fusion: Fusing four antennas is significantly more robust than using any single antenna. Single-antenna models achieved only 70.8–81.5% accuracy, whereas the fused Wi-HAR reached 96.6%. This confirms that different antennas capture uncorrelated "views" of the multipath environment (see Figure 3).
- Uncertainty as a "Zero-Shot" Detector: The system detects environmental shifts without needing examples of those shifts during training. By modeling what "normal" uncertainty looks like for each class, it can identify when the physics of the room has changed.
- Compactness Leads to Interpretability: Because the VAE latent space is so small, the authors could train a shallow decision tree on the latent variables. This tree reached 96.5% accuracy on six activities, providing a human-readable map of how specific antennas contribute to recognizing specific movements (e.g., one antenna might be better at detecting vertical posture changes like "Sitting" vs. "Standing").
Key Results
- Recognition: Reached a 96.6% median accuracy across 12 activities, outperforming state-of-the-art methods like SHARP (70%) on the same benchmark.
- Shift Detection: Achieved a macro-average ROC-AUC of 0.897 for detecting deployment shifts.
- Reliability: At a calibrated 5.2% false-alarm rate, the system successfully rejected 82.5% of samples from shifted environments.
- Interpretability: A depth-four decision tree effectively separated activities using latent variables from multiple antennas, proving the compressed representation retains high-quality physical information.
Limitations and Future Work
- The "What" vs. the "That": While the system is excellent at detecting that a shift occurred, the scalar uncertainty score cannot currently identify what caused the shift (e.g., "new person" vs. "moved receiver").
- Magnitude Only: The current work ignores CSI phase information, which could potentially provide even finer-grained sensing capabilities.
- Adaptation: The authors note that the next logical step is moving from shift detection to shift adaptation—automatically updating the model to handle the new environment once a shift is flagged.
Summary updated