Description
Detecting and classifying intentional, short-duration body or limb motions (waves, swipes, finger pinches) from wireless signals. Gesture recognition is a finer-grained relative of activity-recognition — its motions span tenths of seconds rather than seconds, and target the wrist/hand rather than the torso. It serves as the canonical benchmark for sub-decimeter CSI sensitivity and for cross-domain robustness research.
Why it's hard
- Gestures are short and low-energy; SNR is the binding constraint, especially with commodity NICs.
- Inter-user variability swamps inter-gesture variability — same gesture from two users looks more different than two gestures from one user.
- Tiny location/orientation changes shift the multipath enough to break trained classifiers.
- Multi-person gesture mixing is essentially unsolved in CSI literature.
Common approaches
- BVP / Doppler-spectrogram features fed to CNNs.
- Adversarial domain-adaptation for cross-user / cross-environment robustness.
- Few-shot learning with prototype networks for new-user enrollment.
- Phase-difference engineering across antenna pairs to suppress carrier offset.