Object Detection with YOLO

DelibFlow integrates yolo_ros (v4.5.1) for real-time object detection on TiAGo’s head camera. It runs YOLOv11 by default, but the underlying library supports a bunch of model variants (YOLOv5 through YOLOv12, YOLO-World, YOLOE). There are three nodes involved: one that runs inference, one that tracks objects across frames, and one that overlays detections visually for debugging.

Launching it

The easiest way is through the TiAGo sim launch file:

ros2 launch tiago_sim tiago_yolo.launch.py

This starts YOLOv11 subscribed to head_front_camera/image with a confidence threshold of 0.5, running on cuda:0. The model file (yolo11m.pt) is loaded from the tiago_sim package models directory.

If you want to tweak it, all launch arguments are available:

ros2 launch tiago_sim tiago_yolo.launch.py \
  model:=/path/to/your_model.pt \
  threshold:=0.3 \
  device:=cpu

Launch arguments

Argument	Default	Description
model	yolo11m.pt	Path to the model weights file
tracker	bytetrack.yaml	Tracker config file (bytetrack or botsort)
device	cuda:0	Inference device. Use cpu if no GPU is available
enable	True	Whether to start detection immediately on launch
threshold	0.5	Confidence score threshold for detections
input_image_topic	head_front_camera/image	Image topic to subscribe to
image_reliability	1 (BEST_EFFORT)	QoS reliability (1 = BEST_EFFORT, 2 = RELIABLE)
namespace	""	ROS namespace for all nodes and topics

Yolo Node

This is the main inference node. It is a ROS 2 lifecycle node, so it goes through configure and activate transitions before processing images.

Node name: yolo_node

Topics subscribed

Topic	Interface	Description
image_raw	sensor_msgs/msg/Image	Input image stream from the camera

Topics published

Topic	Interface	Description
detections	yolo_msgs/msg/DetectionArray	Array of all detected objects in the current frame, including bounding boxes, class names, confidence scores, masks, and keypoints

Services

Service	Interface	Description
enable	std_srvs/srv/SetBool	Toggle detection on or off at runtime without restarting the node
set_classes	yolo_msgs/srv/SetClasses	Change which object classes to detect (YOLOWorld models only)

Parameters

Parameter	Type	Default	Description
model_type	string	YOLO	Model variant to load. Options: YOLO, World, YOLOE
model	string	yolov8m.pt	Path to the model weights file
device	string	cuda:0	Compute device (cpu, cuda:0, cuda:1, etc.)
fuse_model	bool	false	Fuse Conv and BN layers for faster inference (not supported by YOLOE)
yolo_encoding	string	bgr8	Image encoding to convert to before passing to the model
enable	bool	true	Start processing images immediately on activation
image_reliability	int	BEST_EFFORT	QoS reliability policy for the image subscription
threshold	double	0.5	Minimum confidence score for a detection to be published
iou	double	0.5	IoU threshold for non-maximum suppression
imgsz_height	int	640	Input image height for inference
imgsz_width	int	640	Input image width for inference
half	bool	false	Use FP16 half-precision inference (faster on supported GPUs)
max_det	int	300	Maximum number of detections per image
augment	bool	false	Apply test-time augmentation for potentially better results at the cost of speed
agnostic_nms	bool	false	Class-agnostic NMS (merges overlapping boxes across different classes)
retina_masks	bool	false	Use full-resolution segmentation masks (higher quality but slower)

Tracking Node

Tracks detected objects across frames using BYTETracker or BOT-SORT, assigning a persistent ID to each object so you can follow it over time.

Node name: tracking_node

Topics subscribed

Topic	Interface	Description
image_raw	sensor_msgs/msg/Image	Input image (time-synced with detections)
detections	yolo_msgs/msg/DetectionArray	Detections from the Yolo Node to associate with tracked IDs

Topics published

Topic	Interface	Description
tracking	yolo_msgs/msg/DetectionArray	Detections annotated with persistent tracking IDs

Parameters

Parameter	Type	Default	Description
tracker	string	bytetrack.yaml	Tracker config file. Use botsort.yaml for BOT-SORT
image_reliability	int	BEST_EFFORT	QoS reliability policy for the image subscription

Debug Node

Overlays detections on the image for visual inspection. Useful for checking that the model is detecting the right things before wiring it into a behavior tree.

Node name: debug_node

Topics subscribed

Topic	Interface	Description
image_raw	sensor_msgs/msg/Image	Input image (time-synced with detections)
detections	yolo_msgs/msg/DetectionArray	Detections to overlay on the image

Topics published

Topic	Interface	Description
dbg_image	sensor_msgs/msg/Image	Image with bounding boxes, masks, and keypoints drawn on it
dgb_bb_markers	visualization_msgs/msg/MarkerArray	3D bounding box markers for RViz
dgb_kp_markers	visualization_msgs/msg/MarkerArray	Keypoint markers for RViz

Parameters

Parameter	Type	Default	Description
image_reliability	int	BEST_EFFORT	QoS reliability policy for the image subscription