YOLOinROS

Compact object detection capability for ROS2. As an object detector, YOLO is implemented via DNN module of OpenCV.

Features

• In regard of Darknet support of OpenCV, this package supports YOLOv2 to v4 and also corresponding Tiny YOLO versions.
• All necessary parameters are stored in a Launch File
• Prediction results are published in two different ways.
  1. In a custom ROS message
  2. In a human readable form as an Image

ROS Topics

Publishers

Topic Name	Message Type
camera/yolo	Image
yolo/array	YoloMsgArray

Subscription

Topic Name	Message Type
camera/image_undist_color	CompressedImage

Getting Started

Dependencies

• OpenCV
  To install:

  $ python3 -m pip install opencv-python

• CvBridge
  To install:

  $ sudo apt install ros-eloquent-cv-bridge

• yolo_msg
  To install:
  Clone repository to /your_workspace/src

  $ git clone https://git.fh-aachen.de/kurhan/yolo_msg.git

  Then, build your workspace

  $ colcon build

• camera_rgb or an alternative image publisher package
  To install:
  Clone repository to /your_workspace/src

  $ git clone https://git.fh-aachen.de/kurhan/camera_rgb.git

  Then, build your workspace

  $ colcon build

Installation

1. Clone repository to /your_workspace/src

   $ git clone https://git.fh-aachen.de/kurhan/yolo_in_ros.git

2. Build /your_workspace with colcon

   $ colcon build

Usage

Launch File

It consist of parameters specific to YOLO

1. config_file contains structure of YOLO and hyperparameters.
   • .cfg file
   • It must be under yolo_in_ros/config/
2. model_file contains weights/parameters obtained during trained.
   • .weight file
   • It must be under yolo_in_ros/config/
3. class_file contains object names that YOLO CNN is trained for.
   • It must be a .txt file.
   • One class per line.
   • It must be under yolo_in_ros/config/
4. conf_thresh_val set the threshold to eliminate predictions with low score.
5. nms_threshold sets the threshold of Non-maximum Supression which eliminates predictions with low confidence score for same objects.
6. width and height set the size of input image of CNN and they must be multiple of 32.
7. scale is a factor which converts 8-bit($[0, 255]$) pixel values to floating numbers in a range of $[0, 1.0]$
8. color indicates that input image has 3 layers(RGB).

• Files of pretrained YOLO can be found here and there
  • For this ROS package the files of Tiny YOLOv4, which is trained by current developers with COCO dataset, are supplied.