yolo_ros_trt¶

Role: ROS 2 wrapper for YOLOv11 instance segmentation, running on TensorRT engines.

Key entrypoints¶

File	Purpose
`yolo_ros_trt/yolo_node.py` (`main`)	`LifecycleNode`. Loads `.engine`, subscribes to image, publishes detections + annotations.
`yolo_ros_trt/tracking_node.py` (`main`)	Optional Ultralytics-tracker wrapper on top of detections.
`yolo_ros_trt/utils/yolo_node_helper.py`	Conversion helpers to `yolo_msgs/DetectionArray`.

ROS interfaces. `yolo_node`¶

Direction	Topic	Type
Sub	`input_image_topic` (default `image`)	`sensor_msgs/Image`
Pub	`output_detections_topic` (default `yolo/detections`)	`yolo_msgs/DetectionArray`
Pub	`output_annotations_topic` (default `yolo/annotations`)	`foxglove_msgs/ImageAnnotations`

Parameters: model_path (required .engine), conf, iou, agnostic_nms, image / detection / annotation topic names, font_size, display_tracker_id.

Models used¶

TensorRT engines from ml_models, e.g.

yolov11s_gate_20250807_0.engine
yolov11s_trash_20250816_0.engine
yolov11s_slalom_20250809_0.engine

Models are stored in .pt form and exported to .engine once per host.

Gotchas¶

Lifecycle

Starts in Inactive. Doesn't publish until activate is called. Confirm with ros2 lifecycle list /yolo_node.

Encoding is BGR8

Inherited from OpenCV / cv_bridge. Anything that consumes annotations should expect BGR.

Tunables require reload

conf / iou apply at NMS time. Changes need a deactivate → activate cycle (or .engine reload) to take effect.

No tracking is built into yolo_node. Use tracking_node if you need it.
Engine generation: Ultralytics.YOLO.export("engine") from a .pt.