ROS YOLOv2 integration

Integrates YOLOv2 with ROS, generating messages with position and confidence of detection.

Requirements

• ROS (tested on Jade and Kinetic)

• GPU supporting CUDA

Installation

If using wstool, the following .rosinstall entry adds this repository to your workspace:

- git:
    local-name: yolo2
    uri: https://github.com/ThundeRatz/ros_yolo2.git

If building from source, change to the src directory of your workspace and use git clone --recursive https://github.com/tiagoshibata/ros_yolo2.git to download with submodules.

This repository is using ROS's nodelet API, which allows loading multiple nodelets in the same process, supporting zero-copy message transfers between nodelets. No node executables are built; instead, shared libraries are. To use them, check the nodelet docs and tutorials. This sample launchfile creates a nodelet manager with an uvc_camera and a yolo2 nodelet:

<launch>
  <group ns="vision">
    <node pkg="nodelet" type="nodelet" name="vision_nodelet" args="manager"/>
    <node pkg="nodelet" type="nodelet" name="camera" args="load uvc_camera/CameraNodelet vision_nodelet" />
    <remap from="yolo2/image" to="image_raw" />
    <node pkg="nodelet" type="nodelet" name="yolo2" args="load yolo2/Yolo2Nodelet vision_nodelet">
      <param name="confidence" value="0.5" />
      <param name="nms" value="0.4" />
    </node>
  </group>
</launch>

Setup

Copy your network config to data/yolo.cfg and network weights as data/yolo.weights.

Message format

For every image received, an ImageDetections message is generated at the yolo2_detections topic, containing a Header and a vector of Detection. The header time stamp is copied from the received image message; this way, a set of detections can be matched to a specific image. Each Detection contains the following attributes:

• uint32 class_id - Class number configured during training;
• float32 confidence - Confidence;
• float32 x - X position of the object's center relative to the image (between 0 = left and 1 = right);
• float32 y - Y position of the object's center relative to the image (between 0 = top and 1 = bottom);
• float32 width - Width of the object relative to the image (between 0 and 1);
• float32 height - Height of the object relative to the image (between 0 and 1).

Contributors

Written by the ThundeRatz robotics team.