ADR-042: Coherent Human Channel Imaging (CHCI) — Beyond WiFi CSI #122
Description
Coherent Human Channel Imaging (CHCI)
WiFi-DensePose currently relies on passive Channel State Information (CSI) extracted from standard 802.11 traffic. CSI is one specific way of estimating a channel response, but it is fundamentally constrained by a protocol designed for throughput and interoperability -- not sensing.
What we actually care about is coherent multipath sensing -- measuring the complex-valued impulse response of the human-occupied channel with sufficient phase stability and micro-Doppler fidelity to reconstruct body surfaces and sub-millimeter physiological motion.
WiFi is optimized for throughput. DensePose is optimized for phase stability. Those goals are not aligned.
CHCI is a purpose-built coherent RF sensing protocol that trades compatibility for control. For sensing, that is a good trade.
What Makes This Different
The Problem with Passive WiFi CSI
Passive CSI sniffing suffers from six fundamental constraints that cannot be fixed in software:
- MAC jitter: CSMA/CA random backoff creates non-uniform sample timing, aliasing Doppler measurements
- Rate adaptation: MCS changes vary bandwidth and modulation between frames -- inconsistent subcarrier count per observation
- LO phase drift: Independent oscillators at TX and RX inject ~5 degrees of phase noise on ESP32, limiting displacement sensitivity to ~0.87 mm at 2.4 GHz
- Frame overhead: 802.11 headers, FCS, and preamble waste airtime that could carry sensing symbols
- Bandwidth fragmentation: Channel bonding decisions by the AP change spectral coverage unpredictably
- Multi-node asynchrony: No shared timing reference means TDM coordination requires statistical phase correction
These constraints impose a hard sensitivity floor. Breathing detection (4-12 mm chest displacement) is reliable. Heartbeat detection (0.2-0.5 mm) is marginal. Body surface reconstruction is limited to volumetric shadows rather than geometric contours.
The CHCI Approach
CHCI replaces passive CSI extraction with intentional, phase-coherent sounding. Six architectural pillars:
1. Intentional OFDM Sounding -- Transmit deterministic NDP (Null Data PPDU) frames at fixed cadence with known pilot structure, compliant with IEEE 802.11bf-2025 (published September 2025). No MAC jitter. No random rate adaptation. No variable bandwidth.
2. Phase-Locked Dual-Radio Architecture -- All nodes share a common reference clock (40 MHz TCXO + SI5351A PLL), distributed via coaxial cable. Both 40 MHz timing and 2.4/5 GHz phase reference signals are distributed. Phase variance drops from ~5 degrees (incoherent) to ~0.5 degrees (coherent). Displacement floor drops from 0.87 mm to 0.031 mm (with 8-antenna averaging).
3. Multi-Band Coherent Fusion -- Simultaneous sounding at 2.4 GHz and 5 GHz (optionally 6 GHz). Lower frequency penetrates walls. Higher frequency increases spatial resolution. Bands are fused contrastively in RuVector embedding space as projections of the same latent motion field, using body model priors to constrain cross-band phase relationships.
4. Time-Coded Micro-Bursts -- Very short (4-20 microsecond) deterministic OFDM bursts at 1-5 kHz cadence. This increases temporal resolution of Doppler shifts without full 802.11 frame overhead. Sensing bandwidth is limited by waveform design, not WiFi framing.
5. MIMO Geometry Optimization -- Antenna spacing tuned for human-scale wavelengths (lambda/4 = 3.125 cm at 2.4 GHz) rather than throughput diversity. L-shaped or linear phased arrays for angular resolution. 4 nodes x 4 antennas = 256 virtual MIMO channels via aperture synthesis.
6. Cognitive Waveform Adaptation -- The waveform adapts in real-time based on scene state. Six sensing modes (IDLE, ALERT, ACTIVE, VITAL, GESTURE, SLEEP) with hysteresis-controlled transitions driven by coherence delta from the body model. RF becomes event-driven. Power consumption drops 60-80% vs constant-rate sounding.
Features
Phase Coherence
- Shared reference clock eliminates per-node LO drift
- Phase variance: 0.5 degrees RMS over 10 minutes (vs 5 degrees incoherent)
- 9x improvement in displacement sensitivity
- Enables reliable heartbeat detection at 2.4 GHz ISM band
Coherent Diffraction Tomography
- Complex-valued channel response (not amplitude-only)
- Body surface reconstruction (geometric contours, not volumetric shadows)
- 3 cm spatial resolution via multi-band synthesis (vs 15 cm single-band)
- Empty-room calibration as reference for channel contrast computation
Sub-Millimeter Vital Signs
- Displacement floor: 0.031 mm at 2 m (8-antenna coherent, 2.4 GHz)
- Breathing rate: target 0.2 BPM error (reference: respiratory belt)
- Heart rate: target 3 BPM error (reference: pulse oximeter)
- HRV (RMSSD) estimation for stress/wellness monitoring
- Apnea detection, fall detection, cardiac anomaly flagging
Cognitive Waveform Engine
- Six sensing modes with automatic transitions
- IDLE: 1 Hz, single band, less than 10 mA radio draw (battery-deployable)
- ACTIVE: 50-200 Hz, all bands, full DensePose + vitals
- VITAL: 100 Hz, optimal subcarrier subset, maximum vital sign fidelity
- Hysteresis prevents rapid mode oscillation (3 consecutive triggers required)
Regulatory Compliance
- IEEE 802.11bf-2025 standard NDP sounding (published standard)
- FCC Part 15.247 compliant (digital modulation, ISM band power limits)
- ETSI EN 300 328 compliant (10 ms burst limit, duty cycle tracking)
- Jurisdiction-aware compliance validator gates all transmissions
Hardware Cost
- $4.25 per CHCI node (ESP32-S3 + antennas + reference clock input)
- $20 total for a 4-node sensing mesh (including reference clock module)
- 10x cheaper than nearest comparable coherent sensing platform
Capabilities
| Capability | Passive WiFi CSI (Current) | CHCI (Proposed) |
|---|---|---|
| Phase coherence | Poor (~5 deg noise) | Excellent (~0.5 deg) |
| Displacement floor | ~0.87 mm | ~0.031 mm (8-antenna) |
| Breathing detection | Reliable | High-precision (0.2 BPM error) |
| Heartbeat detection | Marginal | Reliable |
| Body reconstruction | Volumetric opacity | Surface geometry |
| Spatial resolution | ~15 cm | ~3 cm (multi-band) |
| Doppler resolution | ~30 Hz | ~1 Hz (5 kHz bursts) |
| Through-wall | Excellent | Excellent |
| Power (average) | Constant | 60-80% reduction |
| Cost per node | ~$3 | ~$4.25 |
| Privacy | RF-only | RF-only |
Key Differences from Existing Systems
| System | vs CHCI |
|---|---|
| Passive WiFi CSI | CHCI uses intentional sounding, shared reference clock, coherent phase processing |
| IEEE 802.11bf | CHCI extends 802.11bf with shared-clock coherence, cognitive adaptation, multi-band fusion |
| mmWave Radar | CHCI provides through-wall at 2.4/5 GHz, 10x lower cost, DensePose output vs point cloud |
| UWB (Novelda) | CHCI provides multi-band fusion and DensePose output vs single-band range bins |
| ESPARGOS | CHCI adds DensePose body modeling, cognitive adaptation, multi-band fusion on ESPARGOS-style coherence |
| Google Soli | Soli is 60 GHz close-range gesture only; CHCI is ISM-band room-scale pose + vitals + through-wall |
ADR and DDD Documentation
ADR-042
Full Architecture Decision Record at docs/adr/ADR-042-coherent-human-channel-imaging.md
Six architectural pillars, acceptance tests AT-1 through AT-8, hardware BOM ($4.25/node), 6 implementation phases (~24 weeks), module impact matrix, new crate architecture, 23 references.
DDD Domain Model
Full domain model at docs/ddd/chci-domain-model.md
Six bounded contexts (Waveform Generation, Clock Synchronization, Coherent Signal Processing, Cognitive Waveform, Displacement Measurement, Regulatory Compliance), 16 ubiquitous language terms, Rust struct/enum definitions, domain events, context map with anti-corruption layers.
Acceptance Criteria
Primary: Demonstrate 0.1 mm displacement detection repeatably at 2 meters in a static controlled room.
- AT-1: Phase stability at most 0.5 deg RMS per subcarrier over 10 min static
- AT-2: Displacement detection at most 0.1 mm at 2 m
- AT-3: Breathing rate error at most 0.2 BPM (3 subjects, 5 min each)
- AT-4: Heart rate error at most 3 BPM (3 subjects, seated, 2 min each)
- AT-5: Multi-person pose at least 90% keypoint detection (3 persons, 4x4 m)
- AT-6: IDLE mode power at most 10 mA radio draw
- AT-7: End-to-end pose latency at most 50 ms
- AT-8: Regulatory compliance (FCC 15.247 + ETSI 300 328)
Design Questions
- Should Phase 1 (NDP Sounding) be deployable standalone before full clock distribution hardware?
- Should reference clock use coaxial cable (ESPARGOS-proven) or explore wireless sync (IEEE 1588 PTP)?
- Should cognitive waveform transitions be fully autonomous or require server confirmation?
Related ADRs
ADR-014, ADR-017, ADR-029, ADR-039, ADR-040, ADR-041