Hybrid BLE+CSI Occupancy Estimation

Motivation

Device-based counting (BLE) misses people without phones. Device-free counting (CSI) lacks precision at distinguishing individual occupants. The literature identifies this complementarity but no paper experimentally validates BLE+CSI fusion for crowd counting.

Shahbazian et al. (2023) ↗ dedicates an entire section to "Combined BLE and WiFi" but all referenced works use BLE and WiFi as separate systems, not fused. Park et al. (2022) ↗ uses capture-recapture statistics on device connections but does not combine with CSI. Chaudhari et al. (2024) ↗ discusses BLE + WiFi as "communication-as-sensing" but without experimental fusion.

Setup

Sensing Modalities

Phone Penetration Scenarios

Procedure

Phase 1: Single-Modality Baselines (1 day)

1. Collect data for all 5 phone penetration scenarios (S1-S5)
2. Each scenario: 15 minutes of stable occupancy with manual ground truth
3. Evaluate:
   • BLE-only model: count = number of unique BLE advertisers detected
   • CSI-only model: trained crowd counter from EXP-001/002
   • Record MAE for each modality per scenario

Phase 2: Fusion Model Training (1 day)

Implement and compare 3 fusion approaches:

Approach A — Linear correction: $$\hat{N}{total} = N{CSI} + \alpha \cdot (N_{CSI} - N_{BLE})$$ where $\alpha$ is learned from training data. Intuition: if CSI sees 8 but BLE sees 4, estimate that 50% of people don't have phones, so total is closer to CSI.

Approach B — Learned phone penetration rate: $$\hat{N}{total} = \frac{N{BLE}}{\hat{r}(t)}$$ where $\hat{r}(t)$ is the estimated phone penetration rate, learned from historical data. When CSI data is available, use it to regularize: $\hat{r}(t) = N_{BLE} / N_{CSI}$.

Approach C — Feature-level fusion:

• Input: CSI features (amplitude variance, phase diff) + BLE features (count, avg RSSI, RSSI variance)
• Model: lightweight neural network or gradient boosting
• Output: total occupancy count
• Train on labeled data where manual count is available

Phase 3: Robustness Testing (1 day)

Test fusion model in challenging scenarios:

1. Transient visitors: people walk through the room briefly (BLE detects them, CSI barely registers)
2. Stationary crowd: people sitting still for >30 min (CSI amplitude variance drops per Torun et al. ↗)
3. Mixed activity: half the room moving, half stationary
4. BLE spoofing: one person with multiple BLE devices (tests if fusion handles false positives)
5. Dynamic penetration: people arrive with phones, then put them away (penetration rate changes mid-session)

Expected Outputs

• dataset/exp005/baselines/ — single-modality accuracy per scenario
• dataset/exp005/fusion/ — fusion model accuracy per approach per scenario
• dataset/exp005/robustness/ — challenging scenario results

Analysis Plan

Primary Metrics

1. Fusion accuracy gain: MAE of best fusion model vs. MAE of best single modality
2. Robustness to penetration rate: accuracy vs. phone penetration rate curve
3. Graceful degradation: how does fusion perform as BLE coverage drops to 0%?
4. Upper bound: does fusion at 100% penetration outperform CSI-only? (it should — additional signal)

Key Figure

A line plot with X = phone penetration rate (0-100%), Y = MAE:

• Line 1: BLE-only (linear decrease as penetration drops)
• Line 2: CSI-only (flat — doesn't depend on phones)
• Line 3: Fusion (should be below both lines at all penetration rates)

The intersection of BLE-only and CSI-only lines reveals the "critical penetration rate" — below which CSI is better than BLE. The fusion line should always be below both.

Success Criteria

• Fusion model achieves lower MAE than either single modality at phone penetration rates 25-75%
• At 0% penetration, fusion gracefully falls back to CSI-only accuracy (no degradation)
• At 100% penetration, fusion outperforms CSI-only by leveraging BLE signal
• Feature-level fusion (Approach C) outperforms simple correction models (A, B)

Related Work in Vault

• Human Sensing RF Survey ↗ — "Combined BLE and WiFi" section, no experimental fusion
• CROOD ↗ — capture-recapture on device connections, does not use CSI
• Occupancy Prediction IoT ↗ — BLE for occupancy, no CSI fusion
• Fundamentals Occupancy Detection ↗ — WiFi + BLE as separate sensing, calls for integration
• Zou et al. 2018 ↗ — CSI features for crowd counting, mentions BLE as separate approach
• Torun et al. 2026 ↗ — stationary crowd counting via CSI bandwidth analysis

Dependencies

• EXP-001 (validated BLE ground truth pipeline)
• Trained CSI crowd counting model from EXP-001 or EXP-002
• Participants willing to selectively disable/enable the app during scenarios

Notes

This experiment extends the thesis contribution from "BLE calibrates CSI" (temporal/spatial) to "BLE complements CSI" (modality fusion). It is lower priority than EXP-001 through EXP-003 but strengthens the thesis by showing that BLE is not just a calibration tool but a complementary sensing modality.

The phone penetration rate analysis is particularly relevant for real-world deployment: in a university building, not all students will install the app, so the system must degrade gracefully.