ros2_yolos_cpp

Integrating a YOLO model into a ROS2 robot requires bridging image transport, synchronising message timestamps, managing node lifecycle transitions, and serialising detections into standard ROS2 message types — all before any application logic runs. This plumbing is identical for every robotics project and should not be re-invented.

A complete ROS2 package providing composable, lifecycle-managed nodes for each YOLO task type: detection, segmentation, pose estimation, oriented bounding boxes (OBB), and classification. Every node subscribes to any sensor_msgs/Image topic and publishes results as standard vision_msgs messages.

All configuration is via ROS2 parameters — model path, confidence threshold, NMS threshold, input/output topic names, GPU toggle — so the same node binary handles different models by changing a launch argument. No recompilation needed.

sensor_msgs/Image → CvBridge decode → YOLOs-CPP detector → task-specific publish:

Node	Output topic	Message type
Detector	~/detections	vision_msgs/Detection2DArray
Segmentor	~/detections + ~/masks	Detection2DArray + Image
Pose	~/detections	Detection2DArray with keypoints
OBB	~/detections	ros2_yolos_cpp/OBBDetection2DArray
Classifier	~/classification	vision_msgs/Classification

Lifecycle node design: the ONNX session and GPU context initialise on on_activate and release on on_deactivate. Nodes compose in a single container for efficient resource sharing and are compatible with nav2 lifecycle management. CI/CD tested; strictly typed ROS2 parameters.

• Runs on ROS2 Humble and Jazzy with standard colcon build
• GPU acceleration via use_gpu:=true launch argument (ONNX CUDA execution provider)
• All five task types — detection, segmentation, pose, OBB, classification — from one package
• Docker image provided; runs without local ROS2 install on dev machines
• 148 stars on GitHub; deployed in AMR perception pipelines

CvBridge encoding format mismatches (BGR vs RGB) are the most common integration failure in ROS2 perception stacks. Making the colour conversion explicit in the node — rather than relying on the detector’s internal assumptions — caught encoding bugs that would otherwise surface as silently wrong detections.

Lifecycle nodes add startup complexity but pay back when running multiple models in a container: deactivating one model while keeping another active lets the system adapt to changing task requirements without a full restart.