Physics-Informed Test-Time Adaptation for Mobile HAR
Researchers have rolled out a new framework named PI-TTA aimed at improving human activity recognition (HAR) on mobile devices. This innovative approach addresses challenges in sensor-based HAR, such as variations from sensor rotation, placement shifts, and differences in sampling rates. PI-TTA maintains stability during online updates through three physics-based constraints: ensuring gravity consistency, maintaining short-term temporal continuity, and achieving spectral stability. To avoid issues like overconfident mistakes and representation collapse, it adjusts a small number of parameters. Importantly, the framework allows for on-device personalization with unlabeled test data while ensuring that private information remains decentralized.
Key facts
- PI-TTA is a source-free test-time adaptation framework for HAR on mobile devices.
- It addresses challenges like temporal correlation and within-session shifts.
- Uses three physics-consistent constraints: gravity consistency, temporal continuity, spectral stability.
- Updates a small parameter subset to prevent overconfident errors and catastrophic forgetting.
- Enables on-device personalization from unlabeled test streams without centralizing private data.
Entities
—