Description
Binary detection of whether any moving body is present in the monitored area, irrespective of identity, count, or activity. Motion detection is the lowest-information-content task in wireless human sensing and the easiest one to get right; it is the trigger that wakes up more expensive downstream tasks (crowd-counting, activity-recognition, fall-detection). The thesis treats motion detection as the gating front-end for adaptive duty-cycling of CSI capture.
Why it's hard
- Drift in static-scene CSI baselines causes false positives if simple variance thresholds are used.
- Slow motion (someone reading a book, breathing) sits between the "static" and "moving" regimes.
- Multipath transients from external events (door opening in another room) trigger spurious detections.
- Per-site threshold tuning is required because the static-scene CSI variance differs by orders of magnitude across rooms.
Common approaches
- CSI variance / entropy thresholds with adaptive per-site baselines.
- Subspace or PCA-based change detection over short windows.
- Lightweight binary CNNs for motion vs static classification.
- Coupling with PIR or accelerometer data for confirmation in safety-critical settings.