Vision-Based Precision Landing for Autonomous Drones

Technologies

Python, Raspberry Pi, DroneKit, MAVLink, OpenCV, Computer Vision

Purpose

The goal of this project is to let an autonomous drone land precisely on a defined target - far more accurately than GPS alone allows. The drone navigates to a known GPS position, descends to a set altitude, then uses a downward-facing camera to detect a custom infrared marker and close the remaining distance through closed-loop visual control.

Started in 2022, the project was built on top of a master’s thesis and reworked its detection and control approach end to end. You can access the thesis via avesis.marmara.edu.tr (last accessed: 12 June 2026).

Although the project was started in 2022 and had nearly reached its end by 2024, I only began documenting it in June 2026. Before you make the mistake of asking why it took three years to build and four to document, you can read my blog post - which I’m not proud of.

Features

• Real-time marker detection from a Raspberry Pi camera filtered to 940 nm infrared, isolating the active LED marker from grass, gravel and other natural surfaces.
• Two-stage detection. At higher altitudes the system tracks the centroid of all bright points; once low enough for the full T-pattern to resolve, it switches to geometric pattern recognition.
• Noise-tolerant recognition based on the geometric ratios of the T-pattern, so stray IR reflections get filtered out instead of causing misidentification.
• Closed-loop velocity control. A PID loop closes directly on the marker’s pixel offset from frame center - no per-iteration GPS target computation or coordinate conversion chain.
• Gimbal-mounted camera that stays pointing straight down regardless of how the drone pitches or rolls, removing attitude compensation from the math.
• Flight logging. Every control tick records velocity commands, GPS, altitude and position deltas for post-flight analysis.
• Live video monitoring of the Raspberry Pi screen streamed over LAN during flight.

Usage

The system runs on a Raspberry Pi connected to the flight controller via MAVLink, using DroneKit for vehicle control. Operation proceeds in three stages:

1. Initialize - arm the vehicle, verify sensor state, establish MAVLink communication.
2. Takeoff and search - climb to search altitude and scan for the marker; the drone holds position and rotates until the pattern is found.
3. Navigate and land - compute the marker’s offset from frame center, translate it to a velocity correction, and descend while continuously correcting lateral position.

Altitude is read from LIDAR when in range and the barometer otherwise - used only to gate stage transitions and trigger disarm, not to compute position. The landing pad is a custom T-pattern built from 940 nm infrared LEDs, paired with a matching bandpass filter on the camera.

Limitations

• IR power vs. altitude. Making the LED marker bright enough to be seen from high altitude causes heat buildup and drains the marker’s dedicated battery quickly, which constrains the practical search altitude.
• Low-altitude oscillation. The drone is large relative to the marker, so near the ground the whole vehicle responds to corrections - there’s no isolation between camera and airframe, and wind gusts become control inputs. A horizontally stabilized gimbal loop would decouple this.
• Marker exit at close range. When the drone is directly above and very close, parts of the pattern leave the frame and full-pattern recognition fails. A size-threshold fallback that tracks partial detections is the intended fix.

Special Thanks

• Sateallia - for their support throughout the development process.
• The author and supervisors of the original thesis - for the foundational work this project was built on top of.

Access & License

You can access the project’s source code via github.com/caferozpinar/precision-landing-drone.

Released under the MIT License.