穆再排尔·穆合塔尔
← All work

01 2026 / Innovation project · rated excellent

Autonomous Landing

A three-tier vision pipeline that brings a drone down without GPS.

Role
Project lead
Period
Apr 2025 — Apr 2026
Stack
YOLO26n · RepViT-C3k2 · OpenCV · EPnP · TensorRT · OpenVINO
A red circular landing marker printed on paper, with a central QR code, photographed on a wooden floor.
PL. 01 / Autonomous Landing A red circular landing marker printed on paper, with a central QR code, photographed on a wooden floor. 2026

The drone did not need to be autonomous in the romantic sense. It needed to come down safely on a marker, without GPS, while the camera was moving and the light was bad. The rest was bookkeeping.

Context

GPS-denied landing is a practical problem before it is a research problem. In a field, a roof, or an indoor bay, the signal is absent or untrustworthy. A visual marker gives the vehicle a fixed reference, but only if the pipeline can find it early, track it through motion blur, and recover when the detector loses it.

We collected 1,999 annotated frames across approaches, altitudes, and lighting conditions. The dataset was small enough to audit by hand, which turned out to be the most important decision in the project.

System

The detector is a YOLO26n backbone with a RepViT-C3k2 block, chosen for a better accuracy-per-millisecond trade than the stock model on Jetson. It runs first and provides a coarse box and class.

If the detector is not confident, the pipeline falls back to a QR marker. The QR path is slower but more robust to scale and partial occlusion, and it gives an identity as well as a position. If the QR is also unavailable, the system holds the last known pose and asks the flight controller to hover rather than guess.

The third tier is pose. Once the marker is located, EPnP solves the 3D pose from four coplanar points. The pose is filtered and sent to the flight controller. The important design choice is that each tier can fail independently: detection can fail without losing the marker, QR can fail without losing the pose, and pose can fail without producing a bad command.

The audit

Before training the final model, we built an annotation audit. It checked orientation consistency, agreement between adjacent frames, and a sample of disagreements for human review. It found 367 flipped boxes out of 1,999 — 18.4%.

The detector had been learning the flip. Correcting the labels improved accuracy, but the larger effect was stability: predictions stopped oscillating between two orientations across a sequence. The audit is now part of the training pipeline, not a one-off cleanup.

FIG. 01Three tiers, independent failure paths: detection, QR recovery, and EPnP pose.SWIPE →

Result

The system reaches 85% detection accuracy in complex scenes and runs through TensorRT and OpenVINO on Jetson. It was rated an excellent university innovation project and released as open source.

The number I trust most is not the 85%. It is the 18.4%. The first number describes the model on the data we had. The second describes the data we thought we had. Both matter, and only one of them is usually reported.

A detector cannot be more correct than the sentence it was taught to repeat.

What I learned

The network is the visible part of the work, and the least interesting part to defend. The parts that decided whether the system worked were the fallback logic, the label audit, and the decision to hover when uncertain. Autonomy, in practice, is a careful arrangement of ways to stop.

Measurements

85% complex-scene accuracy
367 flipped labels found
3 fallback tiers
1,999 annotated frames

Next — 02 / 2026

Selective Alignment

What Qwen2.5-VL can decode, and what it refuses to reason about.