UAV / Computer vision / ATAK

MOOSE

Aerial person detection. Computer vision on a fixed-wing UAV, connected to ATAK.

Personal engineering projectPre-flight validation

Concept illustration of UAV computer vision, aerial person detection and transmission to ATAK.

01 / The challenge

The question behind the system.

Bring camera, telemetry, machine perception and ground-station display together on an embedded computer, while keeping the mission software separate from flight-critical control. Aerial images add another difficulty: people can be small, partly hidden and very different from the images a generic detector was trained on.

MOOSE brings together two complementary projects: a Jetson-based mission computer and a dedicated aerial person-detection workflow. Camera frames and flight-controller telemetry feed a ROS 2 pipeline that produces detections, estimated ground positions, live video and mission logs. ArduPilot retains all flight control; the Jetson observes and processes data without commanding aircraft movement.

02 / The approach

From architecture to implementation.

01
Separate the onboard system into ROS 2 packages for camera capture, MAVLink telemetry, vision, geolocation, ATAK output, streaming, health monitoring, logging and startup.
02
Fine-tune YOLOX-S on aerial imagery with sequence-aware dataset splits and duplicate checks. Compare the model against the COCO baseline on independent frozen test sets.
03
Export to ONNX, verify numerical equivalence with PyTorch and integrate TensorRT inference on the Jetson.
04
Validate the complete path from camera and telemetry to ATAK/CivTAK markers and SRT video, including service startup, reboot recovery and mission recording.

03 / Outcomes

What the work shows.

The documented indoor and bench setup runs the integrated mission pipeline: calibrated camera, telemetry, detection, live video, markers, health reporting and logs.
The pre-flight report records nine ROS 2 packages built and 77 tests with no failures or errors; nine tests were skipped.
The model-export workflow verifies PyTorch/ONNX numerical equivalence. Frozen-test analysis records both improvements and regressions across different aerial image sources.

04 / Scope & next steps

Where the work stands.

Quantitative outdoor target-geolocation accuracy and flight validation remain to be established. Bench validation does not establish aircraft flight readiness.
The detector remains sensitive to domain shift, visibility and very small targets. A higher confidence threshold trades recall for precision; there is no single universal accuracy figure.
The Jetson does not control aircraft motion or flight modes. ArduPilot and the RC system retain those responsibilities.

Built on the Flightory Moose airframe, ArduPilot, ROS 2 and YOLOX. Dataset and model terms are documented in the source repositories.

Field notes / Visual documentation