A wall-mounted hub that sweeps a 60 GHz radar to find people through drywall, smoke and fire — then drops each victim into the responder's own AR view, even with no line of sight.
The problem
First responders entering an unfamiliar building have no good way to know where the people inside are — especially if those people are stationary or unconscious and can't call out. A persistently-installed sensor hub plus a responder phone app can give that information cheaply, even through walls.
How it works
Sensing
A TI IWR6843AOPEVM runs TI's Vital-Signs & People-Tracking binary: a 120° field-of-view, 60 GHz radar scan whose on-board DSP identifies people by their detected vitals and movement — work that survives drywall and smoke.
Hub control loop
- An ESP32-S3 uses USB-host to configure and initialize the radar, then ingests its bounding-box output.
- Compass calibration: the installer sweeps the hub through 360°; the stepper is locked at 0° until that's done.
- The stepper then sweeps in 90° steps (0→360→0). With a 120° FOV that overlap covers the full room.
- ESP-NN classifies each track as active, unconscious/stationary, or phantom; if 2 s pass with no real detection, it steps again.
Frame-of-reference math
Detections start in the radar's frame. Using SolidWorks-derived offsets plus the live stepper angle, each position is transformed into the Qorvo DWM3001CDK (UWB) frame, bundled with the compass heading, and streamed over BLE to the responder app.
Responder localization
The app pairs UWB ranging with ARKit's visual-inertial odometry and solves (Gauss-Jordan) for the hub's pose relative to the phone — then re-expresses every detection in the responder's own frame and renders an on-screen indicator, even far away with an occluded view.
Results & honest limits
Through-wall victim localization, rendered live in the responder's AR frame — enough to take 1st place in Analog Devices' Sensor-Fusion track at StarkHacks. It's a prototype: ranging accuracy degrades with multipath, the sweep cadence trades latency for coverage, and a v2 would add multiple hubs for triangulation.