| | Feature | What It Means |

|---|---------|---------------|

| 🔒 | **Privacy-First** | Tracks human pose using only WiFi signals — no cameras, no video, no images stored |

| ⚡ | **Real-Time** | Analyzes WiFi signals in under 100 microseconds per frame — fast enough for live monitoring |

| 💓 | **Vital Signs** | Detects breathing rate (6-30 breaths/min) and heart rate (40-120 bpm) without any wearable |

| 👥 | **Multi-Person** | Simultaneously tracks up to 10 people, each with independent pose and vitals |

| 🧱 | **Through-Wall** | WiFi passes through walls, furniture, and debris — works where cameras cannot |

| 🚑 | **Disaster Response** | Detects trapped survivors through rubble and classifies injury severity (START triage) |

| 🐳 | **One-Command Setup** | `docker pull ruvnet/wifi-densepose:latest` — live sensing in 30 seconds, no toolchain needed |

| 📦 | **Portable Models** | Trained models package into a single `.rvf` file — runs on edge, cloud, or browser (WASM) |

| 🦀 | **810x Faster** | Complete Rust rewrite: 54,000 frames/sec pipeline, 132 MB Docker image, 542+ tests |

WiFi signals penetrate non-metallic debris (concrete, wood, drywall) where cameras and thermal sensors cannot reach. The WiFi-Mat module (`wifi-densepose-mat`, 139 tests) uses CSI analysis to detect survivors trapped under rubble, classify their condition using the START triage protocol, and estimate their 3D position — giving rescue teams actionable intelligence within seconds of deployment.

| Capability | How It Works | Performance Target |

|------------|-------------|-------------------|

| **Breathing Detection** | Bandpass 0.07-1.0 Hz + Fresnel zone modeling detects chest displacement of 5-10mm at 5 GHz | 4-60 BPM, <500ms latency |

| **Heartbeat Detection** | Micro-Doppler shift extraction from fine-grained CSI phase variation | Via ruvector-temporal-tensor |

| **3D Localization** | Multi-AP triangulation + CSI fingerprint matching + depth estimation through rubble layers | 3-5m penetration |

| **START Triage** | Ensemble classifier votes on breathing + movement + vital stability → P1-P4 priority | <1% false negative |

| **Zone Scanning** | 16+ concurrent scan zones with periodic re-scan and audit logging | Full disaster site |

**Triage classification (START protocol compatible):**

| Status | Color | Detection Criteria | Priority |

|--------|-------|-------------------|----------|

| Immediate | Red | Breathing detected, no movement | P1 |

| Delayed | Yellow | Movement + breathing, stable vitals | P2 |

| Minor | Green | Strong movement, responsive patterns | P3 |

| Deceased | Black | No vitals for >30 min continuous scan | P4 |

**Deployment modes:** portable (single TX/RX handheld), distributed (multiple APs around collapse site), drone-mounted (UAV scanning), vehicle-mounted (mobile command post).

**Safety guarantees:** fail-safe defaults (assume life present on ambiguous signals), redundant multi-algorithm voting, complete audit trail, offline-capable (no network required).

The signal processing layer bridges the gap between raw commodity WiFi hardware output and research-grade sensing accuracy. Each algorithm addresses a specific limitation of naive CSI processing — from hardware-induced phase corruption to environment-dependent multipath interference. All six are implemented in `wifi-densepose-signal/src/` with deterministic tests and no mock data.

| Algorithm | What It Does | Why It Matters | Math | Source |

|-----------|-------------|----------------|------|--------|

| **Conjugate Multiplication** | Multiplies CSI antenna pairs: `H₁[k] × conj(H₂[k])` | Cancels CFO, SFO, and packet detection delay that corrupt raw phase — preserves only environment-caused phase differences | `CSI_ratio[k] = H₁[k] * conj(H₂[k])` | [SpotFi](https://dl.acm.org/doi/10.1145/2789168.2790124) (SIGCOMM 2015) |

| **Hampel Filter** | Replaces outliers using running median ± scaled MAD | Z-score uses mean/std which are corrupted by the very outliers it detects (masking effect). Hampel uses median/MAD, resisting up to 50% contamination | `σ̂ = 1.4826 × MAD` | Standard DSP; WiGest (2015) |

| **Fresnel Zone Model** | Models signal variation from chest displacement crossing Fresnel zone boundaries | Zero-crossing counting fails in multipath-rich environments. Fresnel predicts *where* breathing should appear based on TX-RX-body geometry | `ΔΦ = 2π × 2Δd / λ`, `A = \|sin(ΔΦ/2)\|` | [FarSense](https://dl.acm.org/doi/10.1145/3300061.3345431) (MobiCom 2019) |

| **CSI Spectrogram** | Sliding-window FFT (STFT) per subcarrier → 2D time-frequency matrix | Breathing = 0.2-0.4 Hz band, walking = 1-2 Hz, static = noise. 2D structure enables CNN spatial pattern recognition that 1D features miss | `S[t,f] = \|Σₙ x[n] w[n-t] e^{-j2πfn}\|²` | Standard since 2018 |

| **Subcarrier Selection** | Ranks subcarriers by motion sensitivity (variance ratio) and selects top-K | Not all subcarriers respond to motion — some sit in multipath nulls. Selecting the 10-20 most sensitive improves SNR by 6-10 dB | `sensitivity[k] = var_motion / var_static` | [WiDance](https://dl.acm.org/doi/10.1145/3117811.3117826) (MobiCom 2017) |

| **Body Velocity Profile** | Extracts velocity distribution from Doppler shifts across subcarriers | BVP is domain-independent — same velocity profile regardless of room layout, furniture, or AP placement. Basis for cross-environment recognition | `BVP[v,t] = Σₖ \|STFTₖ[v,t]\|` | [Widar 3.0](https://dl.acm.org/doi/10.1145/3328916) (MobiSys 2019) |

**Processing pipeline order:** Raw CSI → Conjugate multiplication (phase cleaning) → Hampel filter (outlier removal) → Subcarrier selection (top-K) → CSI spectrogram (time-frequency) → Fresnel model (breathing) + BVP (activity)

See [ADR-014](docs/adr/ADR-014-sota-signal-processing.md) for full mathematical derivations.

The RuVector Format (RVF) packages an entire trained model — weights, HNSW indexes, quantization codebooks, SONA adaptation deltas, and WASM inference runtime — into a single self-contained binary file. No external dependencies are needed at deployment time.

**Container structure:**

| **Format** | Segment-based binary, 20+ segment types, CRC32 integrity per segment |

| **Progressive Loading** | **Layer A** (<5ms): manifest + entry points → **Layer B** (100ms-1s): hot weights + adjacency → **Layer C** (seconds): full graph |

| **Signing** | Ed25519 training proofs for verifiable provenance — chain of custody from training data to deployed model |

| **Quantization** | Per-segment temperature-tiered: f32 (full), f16 (half), u8 (int8), binary — with SIMD-accelerated distance computation |

| **CLI** | `--export-rvf` (generate), `--load-rvf` (config), `--save-rvf` (persist), `--model` (inference), `--progressive` (3-layer load) |

| **Size** | Typical: 13 KB (sensing-only) to 50+ MB (full DensePose) |

docker run --rm -v $(pwd):/out ruvnet/wifi-densepose:latest --export-rvf /out/model.rvf

The training pipeline implements 8 phases in pure Rust (7,832 lines, zero external ML dependencies). It trains a graph transformer with cross-attention to map CSI feature matrices to 17 COCO body keypoints and DensePose UV coordinates — following the approach from the CMU "DensePose From WiFi" paper (arXiv:2301.00250).

**Three-tier data strategy:**

| Tier | Method | Purpose |

|------|--------|---------|

| **1. Pre-train** | MM-Fi + Wi-Pose public datasets | Cross-environment generalization (multi-subject, multi-room) |

| **2. Fine-tune** | ESP32 CSI + camera pseudo-labels | Environment-specific multipath adaptation |

| **3. SONA adapt** | Micro-LoRA (rank-4) + EWC++ | Continuous on-device learning without catastrophic forgetting |

**Training pipeline components:**

| Phase | Module | What It Does |

|-------|--------|-------------|

| 1 | `dataset.rs` (850 lines) | MM-Fi `.npy` + Wi-Pose `.mat` loaders, subcarrier resampling (114→56, 30→56), windowing, normalization |

| 2 | `graph_transformer.rs` (855 lines) | COCO BodyGraph (17 kp, 16 edges), AntennaGraph, multi-head CrossAttention, GCN message passing |

| 3 | `trainer.rs` (881 lines) | 6-term composite loss (MSE, cross-entropy, UV regression, temporal consistency, bone length, symmetry), SGD+momentum, cosine+warmup scheduler, PCK/OKS metrics |

| 4 | `sona.rs` (639 lines) | LoRA adapters (A×B delta), EWC++ Fisher regularization, EnvironmentDetector (3-sigma drift detection) |

| 5 | `sparse_inference.rs` (753 lines) | NeuronProfiler hot/cold partitioning, SparseLinear (skip cold rows), INT8/FP16 quantization with <0.01 MSE |

| 6 | `rvf_pipeline.rs` (1,027 lines) | Progressive 3-layer loader, HNSW index, OverlayGraph, `RvfModelBuilder` |

| 7 | `rvf_container.rs` (914 lines) | Binary container format, 6+ segment types, CRC32 integrity |

| 8 | `main.rs` integration | `--train`, `--model`, `--progressive` CLI flags, REST endpoints |

**Best-epoch snapshotting**: the trainer saves the best validation loss weights and restores them before checkpoint/export — prevents exporting overfit final-epoch parameters.

./target/release/sensing-server --train --dataset data/ --epochs 100 --save-rvf model.rvf

See [ADR-023](docs/adr/ADR-023-trained-densepose-model-ruvector-pipeline.md).

<details>

<details>

<details>