Vision-Based Autonomous Precision Landing
A precision-landing system that uses AI to detect a helipad "H" marking and lands the drone exactly on its centre — continuously re-centring during descent so wind/drift cannot throw it off. My first serious international freelance job.
Overview
A precision-landing system I delivered in ~2-3 days for an international client via Upwork/Fiverr — my first serious paid international freelance job. A downward-facing camera on the drone detects the helipad "H" marking in real time, and the system brings the drone to the exact centre of that marking before touching down.
The Problem
It had to land on the exact centre of a helipad whose position is not known in advance (variable), despite wind and drift. Landing on a fixed GPS point isn't enough — the target can shift at the moment of touchdown.
How It Works
I detect the "H" with a custom single-class YOLOv3-tiny model (on OpenCV's DNN module). It picks the largest box (nearest pad), finds its centre and computes the offset from the camera centre. A control loop with a dead-zone threshold issues short body-frame (NED) velocity commands (DroneKit + MAVLink) to close the offset. Once centred enough, it switches to ArduPilot LAND mode; it keeps tracking the "H" throughout the descent, makes corrective manoeuvres if it drifts, and disarms the motors on ground contact.
Outputs & Status
I delivered an end-to-end system running in Gazebo + ArduPilot SITL that performs the whole takeoff–search–centre–autonomous-landing loop (demo video on the cover). Since the architecture is DroneKit/MAVLink, it ports to a real ArduCopter vehicle.
Detect → Centre → Autonomous Land
Camera
Downward camera feed
H Detection
YOLOv3-tiny (OpenCV DNN)
Centre & Offset
Box centre ↔ camera centre
Re-centre
NED velocity (DroneKit/MAVLink)
Autonomous Land
ArduPilot LAND + disarm on contact
Tech Stack
Gallery
Looking for a precision / autonomous landing solution?
I can deliver — in days — a vision-based system that detects the target and lands dead-centre despite wind.
I'll build the same for you
Message the assistant — talk through whether I can build something like this, the approach and a ballpark, and get to a booking in minutes.