Real crowds in real buildings — flow, dwell, device-traffic & CSI counting on S3DIS

Question

Take real public-building geometry (not synthetic), populate it with realistic crowds carrying devices that emit traffic into a ray-traced CSI channel, and ask across several floors: (2) what do the flow & dwell patterns look like, (1) can WiFi count the crowd, and (3) does a counter transfer between floors?

Substrate: S3DIS — a real scanned Stanford building (areas 1/5/6), with real furniture (shelves/tables/columns/sofas) and real room layout. Zero invented geometry.

Enabling tooling (built 2026-07-20)

• monad_knowledge/notebooks/python/osm_clean.py — makes any imported floor sim-ready:
  • walkabilize — synthesize shared navigable openings between adjacent rooms + build NRG (real imports have per-room walls / no doors / no nav layer → crowds can't move).
  • bridge <margin> — dilate room cells to close ~62 cm scan wall-gaps → contiguous walkable core.
  • grid <prefix> <nx> <ny> — room-based ceiling AP experiment; clean, analyze.
• device_counting_reduce.py — device-based and device-free counting fidelity (Goal 1) + cross-space transfer (Goal 3).
• Configs sim/exp_csi_crowd/config/s3dis{1,5,6}-csi.yaml (device, seated→network_traffic hooks) and s3dis{1,5,6}-df.yaml (device-free). Grids csi-s3dis{1,5,6}-grid (1 Tx + 24 room APs).
• The OSM path (osm_import_src.py from 07_osm_indoor.ipynb, env-param) + synthetic assembler (public_spaces_assemble.py) were built but not used for the final runs (see Findings).

Goal 2 — Flow & dwell showcase ✅ (interactive atlas)

3 real floors, office agendas (worker/meeter/through/visitor), tuned agent_radius=0.12. 244 agents, 22,482 WiFi device-traffic emissions, coupled ray-traced CSI, all gate-passed on Metal.

s3dis_area5_heatmap.png

Interactive atlas (heatmap/trajectory toggle per floor): the published Artifact.

Goals 1 & 3 — CSI counting on real walled floors

Device-free runs (fixed Tx→24 AP, body-perturbation): area1 01KY0R44EGJ68J506MF2AHPGX8, area5 01KY0R40KCFG2ZGJ3Q0VXB8SAD, area6 01KY0R4DANRNZZA3FZ0GQK5N28.

fig_device_counting.png

• Goal 1 (within-space counting R²): area1 0.64, area5 0.05, area6 0.01 — weak and floor-dependent. Counting works only where ceiling APs overlap occupied rooms; real walls occlude cross-room RF. Sharp contrast to the open-plan mall (c-mall-archcad) where counting succeeded.
• Goal 3 (cross-space transfer R²): all strongly negative (−3 to −67) — a counter trained on one real floor does not transfer to another. Geometry-specific; on-thesis "fingerprints don't transfer".

Findings / honest caveats

1. No clean large real public floor exists off-the-shelf: synthetic looks fake (rejected); OSM indoor is fragmentary (Wien Hbf largest coherent cluster ~6k m²); S3DIS is coherent + furnished but ~1–2.5k m² office. Chose S3DIS (most-real, zero furniture invention).
2. Device-based CSI is occlusion-killed in walled buildings: only 3 of 7,128 seated-device→AP links reach through walls. The "device traffic" emits (emitters.parquet) but is unheard → Goal 1/3 need device-free.
3. CSI crowd counting is fundamentally occlusion-limited in walled multi-room buildings — the headline scientific result of this campaign. Sim-only; a real AX210/Pi5 walled floor (IP-106/IP-112) would confirm the physics.
4. Walkability bridge is floor-dependent (helped area5, over-dilated area1/6) — needs per-floor tuning or true-gap opening placement.

Changelog