Most explanations of GPS-denied navigation stop at the headline: the aircraft works out where it is without satellites. That is true, but it hides the interesting part. Terrain-referenced navigation is an old idea from crewed aviation that has become practical on small aircraft because sensors and compute have shrunk. The idea fits in one sentence. If you know what the ground looks like everywhere, and you can see the ground beneath you, you can work out where you are by finding the patch of map that matches what you see. Everything else is detail: how to sense the ground, store the map, match quickly, and keep the answer stable when the view is ambiguous. This article walks through that detail without the maths.
What the aircraft senses
A terrain-referenced system needs an observation of the ground. In crewed aircraft this was historically a radar altimeter measuring height above terrain along the flight path. On small fixed-wing UAVs the observation is more often a downward-facing camera, sometimes combined with a barometric altimeter and the aircraft’s own attitude estimate.
A camera gives a rich observation: the shape of a creek line, the edge of a road, the texture of a paddock, the boundary of a stand of trees. The catch is that a camera image changes with the time of day, the season, the weather and the altitude. A system that relies on imagery has to be robust to all of that, which is where most of the engineering effort goes.
Alongside the camera sits an inertial measurement unit (IMU), which measures acceleration and rotation many times a second. On its own an IMU can only tell you how you have moved since the last known position. It cannot tell you where you started, and its errors accumulate.
What it compares against
The reference is a terrain database loaded onto the aircraft before flight. It typically has two layers:
- A digital elevation model (DEM), which is a grid of ground heights. Public DEMs cover most of Australia, and higher-resolution products are available from state and commercial sources.
- Georeferenced imagery, such as aerial or satellite orthophotos, which gives the ground’s appearance tied to known coordinates.
The database only needs to cover the planned operating area with a margin, so storage on a small embedded computer is manageable. What matters more is that the data is accurate in the places the aircraft will actually look, and that it is recent enough that the ground has not changed beyond recognition.
The matching step
Given a current observation and a rough idea of where the aircraft is, the system searches the nearby region of the database for the best match. Conceptually this is like sliding a small photo across a large map until the features line up. In practice the comparison works on extracted features, edges and shapes rather than raw pixels, because those survive changes in lighting and colour better.
The output is a position fix, along with an estimate of how confident the system is in it. Confidence is the crucial part. Over distinctive ground, such as a river bend or a road junction, the match is sharp and the fix is trustworthy. Over repetitive or featureless ground the match is broad and the system should say so rather than pretend.
Why drift is the enemy, and why absolute matters
Relative navigation methods, including visual odometry and dead reckoning from the IMU, work by adding up small movements. Each step carries a small error, and those errors compound. After a long enough flight the estimate and the real position can be far apart, and nothing in the method can tell you by how much. The aircraft is confident and wrong.
Terrain-referenced navigation is different in kind. Each successful match ties the aircraft’s position back to coordinates in the database, which are themselves tied to the earth. Errors do not accumulate, because every fix starts fresh. This is why it is described as absolute rather than relative, and why it can act as a genuine substitute for GPS rather than a way of coasting a little longer after GPS is lost.
SLAM and Kalman filtering in plain language
Two terms come up constantly in this field.
SLAM stands for simultaneous localisation and mapping. A robot with no map moves through unknown space, building a map as it goes and working out its own position within it. The map and the position are estimated together, each improving the other. Classic SLAM is relative: the map it builds has no connection to global coordinates unless something anchors it. A terrain-referenced system borrows the machinery of SLAM but replaces the self-built map with a pre-loaded georeferenced one, which is what gives it absolute positioning.
A Kalman filter is the mathematical tool that blends different sources of information according to how much each should be trusted. Picture it as a running estimate of position and velocity that is nudged by each new measurement. Fast, noisy inputs such as the IMU keep the estimate moving smoothly between fixes. Slower, more authoritative inputs such as a terrain match pull the estimate back onto the map. When a match is confident it pulls hard; when it is uncertain it pulls gently. Most autopilots already run one, usually in an extended form to handle the non-linear behaviour of aircraft, which is why adding a new position source to an existing autopilot is possible at all.
Where TerrainSLAM fits
TerrainSLAM applies this approach on a fixed-wing UAV, matching what a downward camera observes against pre-loaded terrain data to produce an absolute position onboard, with no cloud or ground station in the loop. Its processing is designed to handle the real-world imaging problems described above, such as haze, heat shimmer and changing light, and to work across the altitude range from low inspection runs to higher survey flights. It is designed, built and tested in Australia, with patents pending in Australia and the United States.