Description
Identifying or distinguishing individuals by the way they walk, recovered from wireless signals (WiFi CSI, mmWave radar, BLE Doppler) without cooperative devices. Gait is biomechanically idiosyncratic enough to be a soft biometric; the wireless setting trades the ubiquity of the sensor against far lower SNR than a vision-based gait pipeline. For the thesis, gait sensitivity sets a floor on what individuating information CSI exposes, which is relevant both to capability claims and to privacy obligations.
Why it's hard
- Individuating signals are small relative to environmental drift.
- Multi-person scenes mix gait signatures and current methods cannot un-mix them.
- Walking direction, footwear, fatigue, and load all change the gait signature.
- Privacy concerns are real: a re-identifiable RF gait signature is biometric data.
Common approaches
- Doppler-spectrogram CNNs over walk segments.
- BVP-based representations as gait features.
- Physics-aware temporal-embedding learning to separate gait from environmental drift.
- Cross-domain adaptation specifically for gait-style tasks.