AutoLZ

Elevator pitch

A computer vision pipeline for detecting safe landing zones for autonomous drones. YOLOv5 detection model trained on a custom labelled dataset, integrated with control logic for a proof-of-concept demonstration. End-to-end: dataset collection, labelling, augmentation, training, evaluation, integration.

The problem

Autonomous aerial vehicles need to identify safe places to land: not just "flat" but "flat, clear, not water, not moving, not sloped." Human pilots do this by intuition. For autonomous systems, it's a perception problem: given a downward-facing camera feed, output a landing recommendation the flight controller can act on.

The approach

• Collected and labelled a custom dataset of downward-facing aerial images
• Applied preprocessing and augmentation (rotation, brightness variation, occlusion) to survive the noise of real drone camera feeds
• Trained YOLOv5 for landing-zone detection
• Integrated inference with a simple control-logic layer for the demo: given detected safe regions, output the recommended landing target coordinates

Results

Evaluated with confusion matrices and detection metrics. Iterated on hyperparameters and augmentation strategy with a bias toward recall on the "safe" class specifically, because false positives on unsafe zones are a much worse failure mode than missed detections on safe ones. A drone that refuses to land is fine; a drone that lands on water is not.