Abstract
WiFi sensing technology has shown superiority in smart homes among various sensors for its cost-effective and privacy-preserving merits. It is empowered by Channel State Information (CSI) extracted from WiFi signals and advanced machine learning models to analyze motion patterns. However, many learning-based models suffer from severe environmental dependency. While domain adaptation methods have been proposed, they typically require high-quality, well-segmented CSI samples in new environments, which is impractical for end-users. This paper proposes AutoFi, an automatic WiFi sensing model based on a novel geometric self-supervised learning algorithm. AutoFi utilizes unlabeled, low-quality CSI samples captured randomly to learn environment-specific features and then transfers this knowledge to specific user-defined tasks. Evaluation on real-world IoT setups and public datasets (UT-HAR, Widar) demonstrates that AutoFi achieves state-of-the-art performance in human gait and gesture recognition with minimal calibration.
1. Introduction
WiFi sensing leverages Channel State Information (CSI)—data describing how signals bounce off objects and people—to recognize activities without cameras. While promising for privacy-preserving smart homes, current deep learning models are often "brittle." They suffer from environmental dependency: a model trained in a lab often fails in a living room because the signal propagation changes entirely.
To fix this, researchers usually rely on "domain adaptation," which still requires collecting high-quality, well-segmented data in every new location. AutoFi tackles this "cold-start" problem by learning from the "trash" data (random, unlabeled movements) already present in an environment and then quickly adapting to new gestures defined by the user.
2. The AutoFi Framework
AutoFi introduces a two-stage architecture designed to bridge the gap between laboratory training and real-world deployment.
2.1 Geometric Self-Supervised Learning (GSS)
This module learns robust feature representations from unlabeled, randomly collected CSI samples. Unlike standard contrastive learning, AutoFi's GSS forces the model to maintain the geometric relationship between all samples in a batch. It ensures that if two movements are similar in the physical world, their digital embeddings stay close together. This process essentially "maps" the environment's unique signal characteristics without requiring any manual labels.
2.2 Few-Shot Calibration (FSC)
Once the GSS module has organized the feature space, the user provides 1–3 examples (shots) of a specific gesture. AutoFi uses a Prototypical Network to create a "center" for that gesture in the feature space. New movements are then classified based on their proximity to these centers, allowing for rapid deployment of user-defined sensing tasks.
3. Experimental Evaluation
The researchers evaluated AutoFi using three distinct sources:
- Real-world IoT Setup: Data collected using TP-Link routers (Atheros CSI tool) with 114 subcarriers, involving 8 gesture types and gait data from 14 volunteers.
- UT-HAR: A public human activity recognition dataset using Intel 5300 NICs.
- Widar: A large-scale gesture dataset used to test cross-task transferability.
Key Results
- High Accuracy with Minimal Effort: In real-world tests, AutoFi achieved 83.31% accuracy with only 1-shot and 89.71% with 3-shots for gesture recognition.
- Superior Performance: It consistently outperformed state-of-the-art few-shot methods like CSI-GDAM and ReWiS by 4–6%, particularly in subtle tasks like human identification via gait.
- Real-Time Efficiency: While initial self-supervised training takes about 22 minutes offline, the actual inference takes only 15–22ms, making it ideal for real-time IoT applications.
4. Discussion and Future Work
AutoFi demonstrates that unsegmented, random CSI data contains enough environmental "DNA" to pre-train a powerful feature extractor. By enforcing a geometric structure on the feature space, the model "pre-clusters" data, making subsequent few-shot learning highly accurate.
However, limitations remain. The system's performance is sensitive to signal resolution (performing better with 114 subcarriers than 30) and can be affected by extreme dynamic noise, such as multiple people moving simultaneously. Future work will explore multimodal learning—combining WiFi with vision or audio—to further enhance robustness in complex smart home environments.
Summary updated