Skip to content

RF-Based Casualty Cues from Opportunistic Sensors

A simulation-first toolkit for detection, uncertainty, and geolocation—without medical claims

By Benjamin J. Gilbert (Spectrcyde RF Quantum SCYTHE)

TL;DR. We built a fully reproducible, simulation-driven pipeline that uses opportunistic RF sensors—Wi-Fi CSI, BLE RSSI, and UWB—to detect casualty-relevant motion cues, quantify uncertainty, and estimate location via TDoA. It’s engineered for algorithm development and stress-testing, not diagnosis. Code and paper auto-generate figures, tables, and metrics with a single build. Calibration (temperature scaling) and deep ensembles make probabilities honest; robust detectors (micro-Doppler + hysteresis) keep false alarms in check.

Why this matters

Emergencies often unfold where dedicated medical sensors aren’t present—but commodity radios are. Modern buildings, factories, and campuses are soaked in Wi-Fi, BLE, and growing UWB footprints. If we can safely harvest motion-level cues (e.g., consistent micro-Doppler from movement vs. prolonged stillness), we can help responders triage faster, cue cameras without streaming video, and geolocate priority events. The catch: no overreach—this is not a medical device, and we keep it that way.

What we built

  • Physics-informed simulation of Wi-Fi CSI, BLE RSSI, and UWB CIRs across thousands of synthetic scenarios (layout, occlusion, motion regimes).
  • Robust detectors combining robust z-scores + hysteresis with a micro-Doppler energy track (0.3–2 Hz band).
  • A tiny 1D CNN ensemble (ResNet-style) with focal loss (for class imbalance) and temperature scaling (for calibrated probabilities).
  • A ZeroMQ hub that fuses multi-station event onsets into a TDoA heatmap and publishes live geolocation updates.
  • A one-command build that auto-generates all figures, tables, and LaTeX for the paper.

🧯 Scope & ethics: This is a simulation-based development toolkit, not a clinical system. It does not detect blood or diagnose anything. The goal is to benchmark algorithms and quantify uncertainty—safely—before any real-world data collection.

How it works (high-level)

1) Signals → features

  • Wi-Fi CSI: phase-based micro-Doppler spectrograms + subcarrier coherence.
  • BLE RSSI: robust trend/slope features with MAD-based z-scores.
  • UWB CIR: path-energy dynamics and delay-spread changes.

2) Two complementary detectors

  • Rule-based: robust z-score + hysteresis + min-duration gating (crushes flicker).
  • Learned baseline: 1D ResNet-style CNN ensemble over short windows; focal loss (α≈0.25, γ≈2) handles class imbalance; temperature scaling yields calibrated probabilities (good ECE/Brier).

3) Calibrated uncertainty

We calibrate on a held-out split, report ECE/Brier, and select thresholds from PR-optimal operating points. Deep ensembles expose epistemic uncertainty that you can use to down-weight sketchy stations.

4) Geolocation: TDoA

When three or more stations report event onsets, the hub performs a fast grid search in a local ENU frame to produce a TDoA heatmap and best point. Outputs stream on PUB/SUB so dashboards can update in real time.

What we see in simulation

  • The micro-Doppler feature lifts F1 by about ~10–12% over plain energy detectors in motion-heavy scenarios.
  • Calibration holds: ensembles + temperature scaling deliver low ECE and sensible reliability curves.
  • Latency is dominated by window length and hysteresis, not neural compute (forward pass is ~sub-millisecond; wall-clock detection latency typically reflects 1–3 s aggregation).
  • Geolocation accuracy is timing-bound: ~300 m per 1 ms sync error—so GPSDO/NTP discipline matters.

Important: All metrics reflect synthetic validation. They show how the stack behaves, where it breaks, and how to tune thresholds—not field performance.

Figures

(Swap in your generated assets; filenames match the build.)

  • Micro-Doppler waterfall
    figures/micro_doppler.png
    Clear separation between stillness and motion bands (0.3–2 Hz).
  • UWB CIR evolution
    figures/uwb_waterfall.png
    Delay-spread/energy shifts during movement and occlusion.
  • Precision–Recall (ensemble)
    figures/pr_curve.png
    Operating point chosen by PR-optimal threshold (τ*).
  • TDoA heatmap
    figures/tdoa_live.png
    Log-error surface with best point in a ±6 km ENU window.

Reproducibility (one command)

Everything—figures, tables, metrics, and PDFs—builds from source.

# create environment (example)
conda env create -f env.yml
conda activate blood_env

# generate figures + metrics + LaTeX tables
make all

# optional: run the live geolocation demo
make geo-hub         # terminal 1 (starts ZeroMQ hub)
make geo-demo        # terminal 2 (sends 3-station test)
# -> figures/tdoa_live.png and metrics/tdoa_last.json appear

Where this goes next

  1. Data realism. Plug in limited, controlled real captures (with IRB/ethics) to validate timing and false-alarm behavior.
  2. Sensor diversity. Add door sensors, simple acoustic footfall, or mmWave FMCW (public datasets exist) for harder fusion tests.
  3. Outlier-robust geolocation. Weight TDoA by (1–UQ) and add RANSAC over station subsets to survive bad clocks and adversarial noise.
  4. Edge deployment. The 1D path is light enough for phones and small gateways. We’ll benchmark ARM-class devices next.

FAQ

Does this detect “bloodshed”?
No. It detects motion-level cues (e.g., prolonged stillness vs. activity) from RF side-channels. It’s explicitly non-medical and simulation-only in this release.

What about privacy?
We avoid cameras and process low-resolution RF summaries. The pipeline is designed to operate on local gateways; only event metadata needs to leave the site.

Can I use my own floorplans and APs?
Yes—drop them into the simulator, regenerate, and the same build produces new figures/tables.

Want to collaborate?

We’re looking for partners (public safety, campus ops, industrial safety) who can provide timing-disciplined multi-station data in controlled exercises. If that’s you, let’s talk.

(Optional) Front-matter for static sites

Hugo

---
title: "RF-Based Casualty Cues from Opportunistic Sensors"
subtitle: "Simulation-first detection, uncertainty, and TDoA geolocation"
date: 2025-09-07
author: "Benjamin J. Gilbert"
tags: ["RF sensing","Wi-Fi CSI","BLE","UWB","uncertainty","TDoA","simulation"]
images: ["figures/micro_doppler.png"]
draft: false
---

Jekyll

---
layout: post
title: "RF-Based Casualty Cues from Opportunistic Sensors"
description: "A reproducible, simulation-first pipeline for RF motion cues, calibrated uncertainty, and TDoA geolocation."
date: 2025-09-07
author: Benjamin J. Gilbert
tags: [rf, wifi, ble, uwb, uncertainty, geolocation]
image: /figures/micro_doppler.png
---

Social copy (pick one)

  • X/𝕏:
    “New post: a simulation-first pipeline for opportunistic RF sensing (Wi-Fi/BLE/UWB) that detects motion-level casualty cues, quantifies uncertainty, and geolocates via TDoA—no medical claims, just reproducible tooling. 🚑📶🧪”
  • LinkedIn:
    “We released a reproducible RF sensing toolkit that uses Wi-Fi CSI / BLE / UWB to detect motion-level casualty cues and estimate location with TDoA. Calibrated uncertainty, deep ensembles, and a one-command build make it reviewer-friendly and safe. Learn more 👇”

RF-Based Casualty Cues from Opportunistic Sensors – Benjamin J Gilbert CoM EduDownload