Landing is the most critical phase of any fixed-wing flight. For manned aircraft, pilots rely on a combination of visual cues, instrument landing systems (ILS), and ground-based navigation aids. But for autonomous UAVs operating in austere environments — remote airstrips, forward operating bases, or improvised runways — these ground-based aids often don’t exist.
This is where vision-guided landing comes in. By using onboard cameras and real-time semantic segmentation, an autonomous UAV can detect, classify, and align with a runway using nothing but its own eyes.
The Problem
Traditional autonomous landing approaches rely on GPS waypoints and pre-programmed glide slopes. This works well on a calm day at a known airfield, but breaks down in real-world conditions:
- GPS accuracy — standard GPS provides 2-5 meter accuracy. For a UAV with a 3-meter wingspan landing on a 15-meter wide runway, that margin is dangerously thin.
- Unknown or damaged runways — the runway may have obstacles, surface damage, or dimensions that differ from the mission plan.
- GPS-denied environments — electronic warfare, jamming, or simply operating in areas with poor satellite coverage can make GPS unreliable or unavailable.
- Crosswind alignment — GPS tells you where you are, not what the runway looks like ahead. Visual alignment is essential for crosswind corrections.
A vision-based system solves all of these by directly perceiving the runway in real time.
Semantic Segmentation for Runway Detection
At the core of our approach is a lightweight semantic segmentation model running on the UAV’s embedded compute platform. The model classifies every pixel in the forward-facing camera image into categories:
- Runway surface — paved, gravel, or grass landing strips
- Runway markings — centerline, threshold, and touchdown zone markings
- Surrounding terrain — grass, dirt, water, structures, and obstacles
- Sky — used for horizon reference and attitude validation
By segmenting the entire scene at frame rate, the system builds a continuous, pixel-level understanding of the landing environment — far richer than a bounding box detector could provide.
From Pixels to Flight Commands
Raw segmentation output is just a colored image. The real engineering challenge is converting that into actionable flight commands:
Runway geometry extraction. From the segmented runway mask, we compute the runway centerline, width, length, and orientation in the image frame. Using the camera’s known intrinsics and the UAV’s altitude from the barometric altimeter, we transform these pixel measurements into real-world coordinates.
Glide slope computation. The detected runway threshold position, combined with the UAV’s current altitude and distance, defines the required glide slope angle. Our controller continuously adjusts pitch to maintain the target glide slope — typically 3 degrees for a standard approach.
Lateral alignment. The offset between the runway centerline and the image center tells the controller how much lateral correction is needed. This is especially critical in crosswind conditions, where the UAV must crab into the wind while keeping its ground track aligned with the runway.
Flare and touchdown. In the final seconds before touchdown, the system transitions from glide slope tracking to a flare maneuver — reducing descent rate and pitching up slightly to achieve a smooth touchdown. The segmentation model’s detection of the runway threshold triggers this transition.
Why Segmentation Over Detection
A common question: why not just use an object detector to find the runway with a bounding box? There are several reasons:
- Shape matters. A bounding box tells you where the runway is, but not its orientation, width, or centerline position. Segmentation gives you the exact shape.
- Partial visibility. On a long final approach, only part of the runway may be visible. Segmentation handles partial views naturally; detectors struggle.
- Surface condition. Segmentation can distinguish between usable runway surface and damaged or obstructed areas. A bounding box cannot.
- Sub-pixel precision. For landing, you need angular accuracy in the hundredths of degrees. Pixel-level segmentation provides this; bounding boxes do not.
Running at the Edge
Our segmentation model is optimized for real-time inference on embedded hardware:
- Architecture — a lightweight encoder-decoder network designed for the segmentation task, not a repurposed ImageNet backbone
- Resolution — we run at the native camera resolution to preserve the fine details needed for centerline extraction
- Latency — inference completes in under 15 milliseconds on NVIDIA Jetson-class hardware, well within the control loop timing budget
- Robustness — trained on diverse lighting conditions, weather, and runway types to handle real-world variability
Beyond Runways
The same vision-guided approach extends to other precision landing scenarios:
- Ship deck landing — detecting and tracking a moving landing pad on a vessel
- Rooftop landing — identifying safe landing zones on building rooftops for urban UAV operations
- Field landing — selecting and aligning with suitable emergency landing sites in open terrain
- Planetary landing — terrain-relative navigation for spacecraft and Mars/Lunar landers
If you’re developing an autonomous fixed-wing platform and need reliable vision-guided landing, contact us to discuss how our perception and flight control systems can integrate with your vehicle.