ML for Drone Delivery

## From Drone Footage to Autonomous Landing: Building an Aerial Segmentation Model
How does a drone learn to recognize a safe place to land?

In this talk, we’ll walk through the complete development pipeline for an AI-powered drone delivery prototype—from collecting real aerial footage to deploying a computer vision model on PX4 hardware.

You’ll see how drone video is converted into an annotated dataset, how polygon segmentation masks are created in Roboflow, and how a YOLOv11 segmentation model is trained and evaluated in Google Colab.

We’ll then export the trained model to TensorFlow Lite, integrate it into a MAVSDK Android application, and use the model to identify driveways, roads, vehicles, people, sidewalks, yards, and other obstacles in real time.

The session will also cover the practical challenges of class ordering, model outputs, landing-zone logic, waypoint missions, and real-world flight testing. By the end, attendees will understand the full workflow required to move a computer vision model from raw drone imagery to an onboard system capable of locating and centering over a potential delivery zone.

Topics include: YOLOv11, instance segmentation, Roboflow, Google Colab, TensorFlow Lite, MAVSDK SDK, aerial object detection, and autonomous landing-zone selection.