BLE-Assisted CSI Ground Truth Collection

Motivation

Every CSI crowd counting paper in the vault (Wi-CaL ↗, Zou et al. 2018 ↗, Golam Mowla 2024 ↗, EasyCount ↗, Torun et al. 2026 ↗) relies on manual occupancy annotation or scripted scenarios. This makes data collection expensive, error-prone, and difficult to repeat over long durations.

The mobile app can broadcast BLE advertisements at a known interval. A central BLE listener counts unique advertisement sources = automated people count. This provides continuous, timestamped ground truth that can run for hours or days without human intervention.

Setup

Device Deployment

1. CSI TX/RX pairs: Deploy 2-4 ESP32-C5 or Raspberry Pi pairs in the room
   • Arrange links to cross the room diagonally and horizontally
   • Ensure occupants are outside the First Fresnel Zone (FFZ) per Wang et al. 2024 ↗: $F_1 = \sqrt{cD/f}$, where $D$ is TX-RX distance
   • CSI packet rate: 100 Hz (matching Wi-CaL ↗ setup)
2. BLE listeners: 1-2 Raspberry Pi sniffers configured in BLE scan mode
   • Scan window: continuous
   • Filter: only count advertisements matching the app's service UUID
3. Mobile app on participants' phones:
   • BLE advertisement interval: 1 second (BLE 5.0 extended advertising)
   • Unique device identifier in advertisement payload
   • TCP/UDP reporting to collection server (heartbeat every 10s)

Room Configuration

• Room dimensions: measure and record (needed for Fresnel zone calculations)
• Furniture layout: photograph and document
• TX/RX positions: record GPS-free coordinates (tape measure from room corners)
• Door/window state: record (affects multipath)

Data Collection Architecture

Mobile App (BLE TX) ──BLE Adv──> Raspberry Pi (BLE Scanner)
                                        │
                                        ▼
                                 Collection Server
                                        ▲
                                        │
ESP32-C5 (CSI TX) ───WiFi───> ESP32-C5 (CSI RX) ──TCP──┘

Procedure

Phase A: Controlled Validation (1 day)

1. Empty room baseline: 10 min CSI + BLE recording with 0 occupants
2. Incremental occupancy: add 1 person every 5 minutes, up to 10 people
3. Each person carries a phone with the BLE app running
4. Manual count at each transition as ground truth for BLE accuracy validation
5. Record: CSI data (amplitude + phase, all subcarriers), BLE scan results (MAC, RSSI, timestamp), manual count log

Phase B: Natural Occupancy (3-5 days)

1. Deploy in a real meeting room or study area
2. Run continuously during working hours (8:00-18:00)
3. People enter/leave naturally, all carrying phones with the app
4. BLE listener produces automatic occupancy timeline
5. CSI devices record continuously
6. Spot-check manual counts 3x daily to validate BLE accuracy

Phase C: Stress Testing (1 day)

1. Rapid entry/exit: 5 people enter simultaneously, leave simultaneously
2. Partial phone coverage: 3 of 5 people have the app (test undercounting)
3. BLE interference: test with other BLE devices in range (smartwatches, headphones)
4. Edge cases: person stands directly in TX-RX line-of-sight

Expected Outputs

• dataset/exp001/csi/ — raw CSI captures (amplitude + phase per subcarrier per packet)
• dataset/exp001/ble/ — BLE scan logs (timestamp, MAC/UUID, RSSI)
• dataset/exp001/ground_truth/ — manual count log + BLE-derived count timeline
• Synchronization: all devices NTP-synced or PTP-synced to <100ms accuracy

Success Criteria

• BLE-derived occupancy count matches manual count within +/-1 person for >90% of time windows
• CSI data quality: >95% packet capture rate, no hardware dropouts
• Continuous recording: >8 hours without intervention
• Data synchronized across all sources to <500ms

Related Work in Vault

• BLE Can See ↗ — BLE occupancy detection with nRF52840, 95% accuracy. Shows BLE alone is viable for ground truth.
• Human Sensing RF Survey ↗ — comprehensive survey of BLE + WiFi occupancy methods
• Zou et al. 2018 ↗ — CSI features for crowd counting in smart buildings
• Wi-CaL ↗ — ESP32 CSI crowd counting with 4 link pairs, our closest methodological reference
• Zhen et al. 2022 ↗ — adaptive BLE beacon deployment optimization for indoor crowd monitoring

Dependencies

• Mobile app must support BLE extended advertising with configurable UUID
• Collection server (can be Raspberry Pi or laptop) with TCP listener for CSI data and BLE log aggregation
• NTP synchronization across all devices

Notes

This experiment is the foundation for all subsequent experiments (EXP-002 through EXP-005). The automated ground truth pipeline must be validated before it can be trusted for model training and evaluation.