Flying and seeing without a link
The aircraft is meant to launch from a hand throw, fly its planned route, recognise what it sees, save geo-tagged photos and land itself, with no radio link to the ground at all, and to land safely if GPS is lost or jammed. Everything is prepared in our own laptop program, the Ground Station. None of this has flown yet: so far it has been tested in software and in ArduPilot's flight simulator.
We plan in the Ground Station
Our laptop program asks a plain question above every value ("What is the center of gravity of your aircraft?"), and we draw the survey area, launch point and landing strip on a map. It builds the flight controller's settings, the mission and the camera settings from the answers, and checks them all before anything is sent.
Sent by cable, Wi-Fi or USB stick
Settings and mission go to the Matek H743 over USB and are read back to make sure they arrived. The Raspberry Pi 4 gets its settings over its own Wi-Fi hotspot, which switches off while the aircraft is armed, or from a USB stick.
The Raspberry Pi watches the camera
It reads position and attitude from the flight controller, looks for objects in each frame and saves a boxed photo with the object's ground coordinates. An object counts only after it is seen in two frames at the same spot, which removes most one-frame false alarms. It never steers the aircraft.
It lands itself
There is no radio control, so every mission ends with an automatic landing into the wind, and the motor switches off by itself. Low battery, GPS loss or leaving the geofence all lead to the same landing. Above 10 m/s of wind the Ground Station will not let it launch.
Every picture on a map
After landing, the flight comes back to the laptop over Wi-Fi or the USB stick, and the Ground Station shows every picture on a map with its coordinates, filters by object type and exports the list for other map programs.
The Pi watches its own health
It logs its temperature, power supply, free memory and storage, and stops saving ordinary pictures before the storage fills up. If the camera is slow to start or one frame is damaged, it carries on instead of stopping.
GPS drops out or is jammed
Without GPS the aircraft no longer knows exactly where it is.
GPS stays bad for 5 seconds
Carrying on outward would only add to the drift.
A jammer switches off, or sends a false signal
A weak or faked signal can look like a valid fix.
The estimate drifts too far
Dead-reckoning error grows with time and wind.
It lands while GPS is still gone
Without GPS the aircraft can seem to keep moving on the ground and stay armed.
This is a controlled recovery, not navigation without GPS: the longer GPS is gone, the less precise the landing. In the simulator, landings without GPS touched down 85 m past the planned point in 8 m/s of wind and 107 m in 10 m/s, so the landing strip needs at least 150 m of clear ground beyond the touchdown point. At 12 m/s it overshot by 151 m, which is why the Ground Station refuses to launch above 10 m/s.
The detector runs on the Raspberry Pi 4, on the central half of the camera image in four tiles, where the 160° lens is sharpest. Size is the hard limit: at 100 m a car is about 8 pixels for the detector (too few), a combine about 46. The Ground Station works out the highest altitude at which the smallest object you are looking for is still big enough, about 80 m for cars. A colour check finds fire and smoke without any training. The system is set up for 15 classes for farm, forestry and wildfire survey:
| Group | Classes |
|---|---|
| Farm machinery | Tractor, combine, harvester, baler |
| Vehicles | Car, truck, logging truck |
| Farm and forest | Hay bale, livestock, silo, barn, timber stack |
| Hazards | Fire, smoke |
| People | Person |
Our ready-made model knows tractors, cars, trucks and livestock. It was trained on about 1,800 real photos from Google Open Images, then on drone pictures of cattle and sheep taken from above. In drone pictures it had never seen, it found 8 to 9 in 10 of the cattle and sheep (before the drone pictures: 1 to 3 in 10), and 6 to 8 in 10 of its detections were right. The standard model it replaces cannot see tractors at all. New object types can be taught in the Ground Station: add pictures, draw a box around each object, train, and send the model to the aircraft.
Honest status: tractors, cars and trucks were learned from photos taken on the ground, and from straight above, as our camera sees, the model finds few of them; we found no public top-down pictures of them that we may use. Combines are not in it yet, because the public photos of them were unusable. The plan is to fly, save ordinary pictures, label them in the Ground Station, retrain and repeat.
| Check | Result |
|---|---|
| No-RC and failsafe parameters against the ArduPlane source | Every name and range valid |
| GPS-loss script: more than 40 scenarios, including jamming, takeoff, landing and never switching off in flight | All pass |
| Full missions with GPS jamming in ArduPilot's simulator (SITL), wind 4 to 10 m/s, some with GPS gone until after the landing | 6 of 6 pass |
| The same at 12 m/s of wind | Lands 151 m long, so 10 m/s is our limit |
| A fresh simulated flight controller set up only by the Ground Station, then flown with GPS jamming | Passes |
| Ground Station used like a person would, in a real browser: every tab, the map, sending, the launch checklist, the pictures | Passes |
| Recognition model on 298 photos it had never seen | Precision 0.70 to 0.75 |
| Whole onboard pipeline, from camera frame to saved, geo-tagged photo; camera that is late or drops frames | Passes |
| Flight tests | After the airframe is built |
400 automated tests run on every change. They tell us the software does what we intended; only flight tests will tell us whether that holds up in the air.
SILM Buteo is in development and has not flown yet. Figures are design estimates from CFD and calculation, not flight-test data. About the student project