Adds WindowedMlpModel: 440 → 64 ReLU → n_classes, stacks last 20
frames × 22 features as input. Captures temporal patterns that
frame-level classifiers physically cannot see (walking cadence,
sit-stand cycles, gesture rhythm).
AppStateInner gets feature_window: VecDeque<[f64; 22]> (cap 20)
auto-pushed at the 3 tick sites before adaptive_override. The
classify_window API flattens the buffer (oldest first) + current
frame's features → 440-d input → softmax over classes. Cold-start
(<20 frames) falls back to frame-level MLP.
AdaptiveModel now carries all three classifiers side-by-side:
LogReg (ADR-118), MLP (ADR-119), W-MLP (this). classify_window
picks W-MLP first; legacy classify() picks MLP > LogReg.
Result on the same 6-node, 7-class, 151,329-frame dataset:
LogReg: 49.58%
MLP: 53.53%
W-MLP: 90.40% (+36.87 pts over MLP, +50.0 pts over original
2-node 15-feature LogReg baseline)
Per-class W-MLP accuracy:
absent 100% (was 41%)
present_still 100% (was 99%, saturated)
transition 86% (was 36%) — sit/stand cadence captured
waving 90% (was 38%) — gesture cadence captured
present_moving 82% (was 33%) — walking step cadence captured
active 74% (was 30%) — jumping bursts captured
Loss broke through frame-level plateau (1.15 → 0.25). Caveat:
90.4% is training-set accuracy; ~28k weights on ~30k windowed
samples means some overfitting likely. Held-out test set
recommended as follow-up.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
|
||
|---|---|---|
| .. | ||
| src | ||
| static | ||
| tests | ||
| Cargo.toml | ||
| README.md | ||
README.md
wifi-densepose-sensing-server
Lightweight Axum server for real-time WiFi sensing with RuVector signal processing.
Overview
wifi-densepose-sensing-server is the operational backend for WiFi-DensePose. It receives raw CSI
frames from ESP32 hardware over UDP, runs them through the RuVector-powered signal processing
pipeline, and broadcasts processed sensing updates to browser clients via WebSocket. A built-in
static file server hosts the sensing UI on the same port.
The crate ships both a library (wifi_densepose_sensing_server) exposing the training and inference
modules, and a binary (sensing-server) that starts the full server stack.
Integrates wifi-densepose-wifiscan for multi-BSSID WiFi scanning per ADR-022 Phase 3.
Features
- UDP CSI ingestion -- Receives ESP32 CSI frames on port 5005 and parses them into the internal
CsiFramerepresentation. - Vital sign detection -- Pure-Rust FFT-based breathing rate (0.1--0.5 Hz) and heart rate (0.67--2.0 Hz) estimation from CSI amplitude time series (ADR-021).
- RVF container -- Standalone binary container format for packaging model weights, metadata, and
configuration into a single
.rvffile with 64-byte aligned segments. - RVF pipeline -- Progressive model loading with streaming segment decoding.
- Graph Transformer -- Cross-attention bottleneck between antenna-space CSI features and the
COCO 17-keypoint body graph, followed by GCN message passing (ADR-023 Phase 2). Pure
std, no ML dependencies. - SONA adaptation -- LoRA + EWC++ online adaptation for environment drift without catastrophic forgetting (ADR-023 Phase 5).
- Contrastive CSI embeddings -- Self-supervised SimCLR-style pretraining with InfoNCE loss, projection head, fingerprint indexing, and cross-modal pose alignment (ADR-024).
- Sparse inference -- Activation profiling, sparse matrix-vector multiply, INT8/FP16 quantization, and a full sparse inference engine for edge deployment (ADR-023 Phase 6).
- Dataset pipeline -- Training dataset loading and batching.
- Multi-BSSID scanning -- Windows
netshintegration for BSSID discovery viawifi-densepose-wifiscan(ADR-022). - WebSocket broadcast -- Real-time sensing updates pushed to all connected clients at
ws://localhost:8765/ws/sensing. - Static file serving -- Hosts the sensing UI on port 8080 with CORS headers.
Modules
| Module | Description |
|---|---|
vital_signs |
Breathing and heart rate extraction via FFT spectral analysis |
rvf_container |
RVF binary format builder and reader |
rvf_pipeline |
Progressive model loading from RVF containers |
graph_transformer |
Graph Transformer + GCN for CSI-to-pose estimation |
trainer |
Training loop orchestration |
dataset |
Training data loading and batching |
sona |
LoRA adapters and EWC++ continual learning |
sparse_inference |
Neuron profiling, sparse matmul, INT8/FP16 quantization |
embedding |
Contrastive CSI embedding model and fingerprint index |
Quick Start
# Build the server
cargo build -p wifi-densepose-sensing-server
# Run with default settings (HTTP :8080, UDP :5005, WS :8765)
cargo run -p wifi-densepose-sensing-server
# Run with custom ports
cargo run -p wifi-densepose-sensing-server -- \
--http-port 9000 \
--udp-port 5005 \
--static-dir ./ui
Using as a library
use wifi_densepose_sensing_server::vital_signs::VitalSignDetector;
// Create a detector with 20 Hz sample rate
let mut detector = VitalSignDetector::new(20.0);
// Feed CSI amplitude samples
for amplitude in csi_amplitudes.iter() {
detector.push_sample(*amplitude);
}
// Extract vital signs
if let Some(vitals) = detector.detect() {
println!("Breathing: {:.1} BPM", vitals.breathing_rate_bpm);
println!("Heart rate: {:.0} BPM", vitals.heart_rate_bpm);
}
Architecture
ESP32 ──UDP:5005──> [ CSI Receiver ]
|
[ Signal Pipeline ]
(vital_signs, graph_transformer, sona)
|
[ WebSocket Broadcast ]
|
Browser <──WS:8765── [ Axum Server :8080 ] ──> Static UI files
Related Crates
| Crate | Role |
|---|---|
wifi-densepose-wifiscan |
Multi-BSSID WiFi scanning (ADR-022) |
wifi-densepose-core |
Shared types and traits |
wifi-densepose-signal |
CSI signal processing algorithms |
wifi-densepose-hardware |
ESP32 hardware interfaces |
wifi-densepose-wasm |
Browser WASM bindings for the sensing UI |
wifi-densepose-train |
Full training pipeline with ruvector |
wifi-densepose-mat |
Disaster detection module |
License
MIT OR Apache-2.0