m monad-knowledge FIIT STU · open research vault
unclassified superseded

Multi-Room Generalization with BLE Transfer

A CSI crowd counting model trained in Room A can be transferred to Room B using only a 10-30 minute BLE calibration campaign, achieving >80% of the accuracy of a model fully trained in Room B.

All experiments

Environment

FIIT STU — 2 rooms with different dimensions, furniture, and multipath characteristics

Motivation

Santos et al. (2024) ↗ found that most CSI crowd counting systems fail across rooms — models trained in one environment perform poorly in another due to different multipath profiles. Guarino et al. (2026) ↗ identifies cross-environmental generalization as a critical unsolved challenge. Chen et al. (2023) ↗ surveys domain adaptation methods but none propose BLE-assisted few-shot adaptation for crowd counting.

This experiment tests whether the BLE calibration mechanism from EXP-003 also enables efficient cross-room transfer — reducing the deployment cost of CSI sensing from "hours of labeled data per room" to "one 10-minute BLE campaign per room."

Setup

Room Selection

Property Room A (source) Room B (target)
Dimensions e.g., 5x5m (meeting room) e.g., 8x10m (seminar room)
Furniture Dense (table, chairs) Sparse (open layout)
TX-RX distance ~4m ~7m
Max occupancy 10 20
Multipath profile Strong reflections (small room) Weaker, more paths (large room)

Choose rooms that are meaningfully different — different sizes, different furniture density, different wall materials if possible. This maximizes the domain gap.

Device Deployment

  • Room A: Full EXP-001 deployment (CSI pairs + BLE listeners)
  • Room B: Identical hardware, but different physical arrangement appropriate for the room
  • Keep TX-RX link geometry similar (diagonal + horizontal) but at different distances

Procedure

Phase 1: Full Training in Room A (1 day)

  1. Collect 2 hours of CSI + BLE ground truth data in Room A (incremental occupancy)
  2. Train 3 models (Random Forest, LSTM, CNN) on Room A data
  3. Evaluate on Room A held-out test set → record as "Room A accuracy" (upper bound)

Phase 2: Zero-Shot Transfer to Room B (1 day)

  1. Deploy Room A models to Room B without any Room B training data
  2. Collect 2 hours of CSI data with BLE ground truth in Room B
  3. Evaluate Room A models on Room B data → record as "zero-shot accuracy" (lower bound)
  4. Expected: significant accuracy drop (per Santos et al. ↗)

Phase 3: BLE-Assisted Transfer (1 day)

  1. Run a BLE calibration campaign in Room B: 10 minutes, 5 participants
  2. Fine-tune Room A models on the 10-minute Room B campaign data
  3. Evaluate fine-tuned models on remaining Room B data → record as "BLE-transfer accuracy"
  4. Repeat with 20-minute and 30-minute campaigns to measure duration impact

Phase 4: Full Training Baseline in Room B (1 day)

  1. Collect 2 hours of full training data in Room B
  2. Train fresh models from scratch → record as "Room B native accuracy" (gold standard)
  3. Compare: BLE-transfer accuracy vs. Room B native accuracy

Phase 5: Bidirectional Transfer (optional)

  1. Train model in Room B
  2. Transfer to Room A with BLE campaign
  3. Check if transfer is symmetric or if one direction is easier

Expected Outputs

  • dataset/exp004/room_a/ — full training data + CSI + BLE
  • dataset/exp004/room_b/ — campaign data + evaluation data + CSI + BLE
  • dataset/exp004/results.csv — accuracy comparison table

Analysis Plan

Key Comparison Table

Scenario Training Data Expected Accuracy
Room A native 2h Room A High (reference)
Zero-shot B 0 Room B Low (problem evidence)
BLE transfer 10min 10min Room B Medium-high?
BLE transfer 30min 30min Room B Higher?
Room B native 2h Room B High (gold standard)

Primary Metrics

  1. Transfer efficiency: BLE-transfer accuracy / Room B native accuracy (target: >80%)
  2. Data efficiency: BLE-transfer accuracy achieved with 10 min vs. Room B native with 120 min = 12x reduction
  3. Accuracy gap: Room B native - BLE transfer (should be <10%)
  4. Minimum campaign duration: shortest campaign that achieves >80% of native accuracy

Success Criteria

  • Zero-shot transfer shows measurable degradation (>15% accuracy drop) — validates the problem
  • 10-minute BLE campaign recovers >50% of the accuracy gap between zero-shot and native
  • 30-minute BLE campaign achieves >80% of Room B native accuracy
  • At least one model architecture transfers well (RF or CNN typically better than LSTM for transfer)
  • Santos et al. 2024 ↗ — "lack of studies investigating environment independence for WiFi CSI CC systems"
  • WiFi Sensing Generalizability ↗ — taxonomy of cross-domain WiFi sensing techniques
  • Cross-Domain WiFi Sensing Survey ↗ — domain-invariant feature extraction methods
  • Meneghello 2023 ↗ — few-shot domain adaptation: "retraining some parameters based on a few examples from the new domain"
  • CSI Sensing Datasets Survey ↗ — "collect data from heterogeneous environments" as a key challenge

Dependencies

  • EXP-001 (validated BLE ground truth)
  • Access to 2 different rooms with power and network connectivity
  • Duplicate set of CSI/BLE devices for Room B (or move devices between phases)

Notes

This experiment extends EXP-003 from temporal transfer (same room, different time) to spatial transfer (different room, same time). Together, EXP-003 + EXP-004 demonstrate that BLE calibration campaigns solve both temporal drift and spatial generalization — the two main failure modes identified in the literature.

The data efficiency metric is the key selling point: "deploying CSI crowd counting in a new room requires 10 minutes instead of 2 hours."

Provenance

—
not recorded

Data types

  • csi-amplitude
  • csi-phase
  • ble-ground-truth
  • transfer-learning-metrics