wifi-densepose/examples/research-sota/04-rssi/README.md

1.5 KiB
Raw Blame History

04 — RSSI-only sensing

RSSI is the simplest CSI summary (one number per packet). These scripts quantify what's recoverable from RSSI alone vs full CSI.

Scripts

Script Thread Headline
r8_rssi_only_count.py R8 RSSI-only person count: 59.1% accuracy = 94.82% of full-CSI v0.0.2 with a tiny 656-parameter MLP. RSSI keeps 95% of counting capacity.
r9_rssi_fingerprint_knn.py R9 Cosine-NN on RSSI fingerprints: 2.18× lift over chance (MODERATE). Surfaces counting-vs-localization asymmetry: RSSI is great for count, weaker for per-location ID.

The counting-vs-localization asymmetry

R8 + R9 together demonstrate that RSSI:

  • Retains 95% of person-count capacity (R8)
  • Retains only ~30% of localization capacity (R9)

This means RSSI-only deployments (the cheap path) are viable for occupancy / count but inadequate for per-occupant features (vitals, identity, pose).

When to use RSSI-only

Per ADR-113 placement matrix, RSSI-only is appropriate for:

  • cog-presence (binary occupancy)
  • cog-person-count (occupant count)
  • Very cost-sensitive deployments (chicken-scale R19 livestock, for instance)

NOT appropriate for:

  • cog-vital-signs (needs CSI per-subcarrier shape)
  • cog-pose-estimation (needs CSI multistatic geometry)
  • cog-quantum-vitals (ADR-114, needs CSI fusion with NV)

See also

  • Research notes: docs/research/sota-2026-05-22/R8-*.md, R9-*.md
  • Composes with: 01-physics-floor/ (uses Fresnel forward model insight)