Skip to content

Object Detection

Summary

This node detects buoys and boats on the water and estimates their bearing (angle) and range (distance).
Detection is powered by an Ultralytics YOLOv11 model. We use the Ultralytics results object to retrieve bounding boxes, classes, and confidences.

The node runs on ROS 2 and publishes: - Bearing to target (angle) — relative to the camera/boat frame. - Range to target (depth distance) — from depth data or geometric approximation. - Per-detection metadata — confidence score, relative x/y of the bounding box, etc.

What it publishes

  • sensor_msgs/msg/Image
  • sailbot_msgs/msg/ObjectDetectionResults
  • detections[]: { confidence value, x_postion, y_position}

QoS is configurable via QoSProfile and should be tuned to match the camera publisher (e.g., RELIABLE vs BEST_EFFORT, KEEP_LAST depth, etc.).


Repository (Models & Dataset)

The code, models, and dataset live on Hugging Face (handles large files better than GitHub):

Object Detection – Hugging Face Repository:
https://huggingface.co/datasets/Aanimated/autoboat_vt_object_detection/tree/main

We use a Hugging Face dataset repo because it supports large images and model weights via Git LFS.
You can clone it and pull all LFS-tracked files locally.

Quick start

# Install Git LFS, clone, and pull large files
sudo apt update && sudo apt install -y git-lfs
git lfs install

git clone https://huggingface.co/datasets/Aanimated/autoboat_vt_object_detection
cd autoboat_vt_object_detection
git lfs pull