Amber Sentinel
See the fire, not the roof.
A thermal camera sees heat. It can't tell a wildfire from a sun-baked roof.
Amber Sentinel adds a second camera that can, so real fires get flagged and hot roofs, cars, and rocks don't.
01 / Thermal camera
02 / Visible camera
03 / Result
One false alarm removed
Illustration of how it works, not real camera footage.
Real hardware
Not just a drawing
This is the turret in hand: the low-cost version, which carries only the visible-light camera.
- About $120
- for the whole camera unit
- 2 axes
- pan and tilt, both servo-driven
The full version adds a FLIR thermal camera to the same turret. Fitting it in is the current step.
How it works
Four steps, one decision
- 01
Thermal finds candidates
A thermal camera scans for hot spots. Anything warm enough becomes a candidate, so small fires don't slip through.
- 02
The visible camera checks each one
A visible-light camera is aimed at exactly the same spot. A detection model looks for flame and smoke there. A hot spot with no fire in view, like a roof, is rejected.
- 03
Confirmed over several frames
A candidate has to hold up across consecutive frames before it counts, which filters out one-frame glitches.
- 04
A person decides
Confirmed detections go to a human operator. Amber Sentinel finds fires; people decide what to do about them.
This is the design. Progress below shows which parts are built and tested.
Hardware
Built to move between airframes
- Turret
- Two-axis ball turret, 3D-printed, with servo-driven pan and tilt. Four design versions so far.
- Thermal camera
- FLIR 640 × 512 thermal core.
- Visible camera
- Sony IMX477, cropped to match the thermal camera's field of view.
- Onboard computer
- NVIDIA Jetson Orin Nano, running the detection models as TensorRT engines. No ground link is needed to detect.
- Detection models
- Custom YOLOv8 models trained on public wildfire datasets plus 938 aerial frames labeled by hand.
- Versions
- Low-cost visible-light only, about $120 for the camera unit. Full version adds the FLIR thermal camera.
- Airframe
- Designed to mount on more than one drone. The first test platform is a fixed-wing airframe.
Progress
Bench-tested, not yet flown
73%
Fire precision at 40% recall, on aerial footage the model never trained on. Target: 70%.
66%
Smoke precision at the same recall. Below target and still in progress.
100–130 ms
Per frame on the Jetson, about 8 to 10 frames per second.
—
Not yet measured with both cameras together. Combining them in the turret is the current step.
- DEC 2025
Project starts.
- 2026
Turret designed and rebuilt through four versions.
- SEP 2026
Final turret mechanism passes its bench test. Detection models benchmarked on the onboard computer.
- OCT 2026
Test airframe's flight controller configured and control surfaces tested.
- NOWIN PROGRESS
Integrating the thermal and visible cameras in the turret.
- NEXTUPCOMING
First flight of the test airframe, without the turret.
- SPRING 2027PLANNED
Collecting flight data over controlled burns.
Who it's for
Wildfire agencies in California and the western US, where large areas need watching and a fire caught in its first minutes is far easier to stop. Local fire departments are a second fit.
Team
- Aiden LiuDesigned and built Amber Sentinel: the turret, electronics, detection models, and flight setup.
- Jason LiuConrad Challenge teammate.
- Ce LiuCoach, and pilot in command for test flights.
Developed by Stallion Aerial Systems and entered in the 2026–27 Conrad Challenge, Aerospace & Aviation.
Contact
Fire departments, researchers, and anyone with wildfire imagery or a controlled burn to fly near: I'd like to hear from you.
contact@ambersentinel.com