Description
Classifying the activity a person is performing from sensor observations — walking, sitting, standing, falling, running, eating, typing — typically at the second to minute timescale. In WiFi-CSI literature HAR is the canonical "is the link doing something useful?" benchmark; it is also the closest neighbor of crowd-counting in algorithmic terms because the same CSI features feed both. The thesis uses HAR-style features as inputs to count regressors and as a sanity check that CSI is detecting bodies rather than environmental noise.
Why it's hard
- Activity classes overlap in their CSI signatures (sitting vs standing-still are close to indistinguishable).
- Cross-environment transfer is poor (see environment-dependence).
- Multi-person scenes mix signatures additively with strong interference between subjects.
- Sample efficiency: deep models need thousands of labeled segments per class per site.
- Fine-grained activities (typing, gestures) demand higher CSI resolution than is available with commodity NICs.
Common approaches
- CNN/LSTM/Transformer over CSI amplitude and phase time series.
- Body-Coordinate-Velocity-Profile (BVP) and Doppler-frequency-shift features.
- Self-supervised pretraining on unlabeled CSI plus fine-tuning per environment.
- Multi-modal fusion (CSI + IMU) for robustness.
Source Papers
- wang2015_48cf ↗ — understanding and modeling of WiFi-signal HAR.
- guo2024_9632 ↗ — HAR via WiFi + inertial sensors with ML.
- logah2026_c3bb ↗ — efficient ML for WiFi CSI HAR via Monte Carlo features.
- ahmad2024_8639 ↗ — WiFi-based human sensing with deep learning (review).
- ullmann2023_0ac3 ↗ — radar-based continuous HAR (survey, comparison baseline).