Description
Gesture recognition classifies a short window of CSI (or other RF observations) into a predefined gesture vocabulary — pushes, pulls, draws, pose changes. WiFi gesture pipelines rely heavily on Doppler features from CSI: the spectrogram of velocity components vs time forms a signature that distinguishes gestures. Widar3.0 popularised body-coordinate-velocity-profile (BVP) features that are partially environment-invariant.
When it's used
- Touchless human-computer interaction
- Smart-home control benchmarks (Widar, SignFi)
- Cross-domain generalisation studies (different rooms, users)
- Few-shot / meta-learning evaluations
Limitations
- Vocabulary tied to the training set; open-set recognition is open
- Multi-user gestures interfere on the same Doppler bins
- Environment shift collapses accuracy unless BVP-style domain-invariant features are used