DNN Inference Nodes for ROS/ROS2

This package contains DNN inference nodes and camera/video streaming nodes for ROS/ROS2 with support for NVIDIA Jetson Nano / TX1 / TX2 / Xavier / Orin devices and TensorRT.

The nodes use the image recognition, object detection, and semantic segmentation DNN's from the jetson-inference library and NVIDIA Hello AI World tutorial, which come with several built-in pretrained networks for classification, detection, and segmentation and the ability to load customized user-trained models.

The camera & video streaming nodes support the following input/output interfaces:

• MIPI CSI cameras
• V4L2 cameras
• RTP / RTSP streams
• WebRTC streams
• Videos & Images
• Image sequences
• OpenGL windows

Various distribution of ROS are supported either from source or through containers (including Melodic, Noetic, Foxy, Galactic, Humble, and Iron). The same branch supports both ROS1 and ROS2.

Table of Contents

• Installation
• Testing
  • Video Viewer
  • imagenet Node
  • detectnet Node
  • segnet Node
• Topics & Parameters
  • imagenet Node
  • detectnet Node
  • segnet Node
  • video_source Node
  • video_output Node

Installation

The easiest way to get up and running is by cloning jetson-inference (which ros_deep_learning is a submodule of) and running the pre-built container, which automatically mounts the required model directories and devices:

$ git clone --recursive --depth=1 https://github.com/dusty-nv/jetson-inference
$ cd jetson-inference
$ docker/run.sh --ros=humble  # noetic, foxy, galactic, humble, iron

note: the ros_deep_learning nodes rely on data from the jetson-inference tree for storing models, so clone and mount jetson-inference/data if you're using your own container or source installation method.

The --ros argument to the docker/run.sh script selects the ROS distro to use. They in turn use the ros:$ROS_DISTRO-pytorch container images from jetson-containers, which include jetson-inference and this.

For previous information about building the ros_deep_learning package for an uncontainerized ROS installation, expand the section below (the parts about installing ROS may require adapting for the particular version of ROS/ROS2 that you want to install)

$ cd ~
$ sudo apt-get install git cmake
$ git clone --recursive --depth=1 https://github.com/dusty-nv/jetson-inference
$ cd jetson-inference
$ mkdir build
$ cd build
$ cmake ../
$ make -j$(nproc)
$ sudo make install
$ sudo ldconfig

$ sudo apt-get install ros-melodic-image-transport ros-melodic-vision-msgs

$ sudo apt-get install ros-eloquent-vision-msgs \
                       ros-eloquent-launch-xml \
                       ros-eloquent-launch-yaml \
                       python3-colcon-common-extensions

$ mkdir -p ~/ros2_example_ws/src

$ cd ~/ros_workspace/src
$ git clone https://github.com/dusty-nv/ros_deep_learning

$ cd ~/ros_workspace/

# ROS Melodic
$ catkin_make
$ source devel/setup.bash 

# ROS2 Eloquent
$ colcon build
$ source install/local_setup.bash 

Testing

Before proceeding, if you're using ROS Melodic make sure that roscore is running first:

$ roscore

If you're using ROS2, running the core service is no longer required.

Video Viewer

First, it's recommended to test that you can stream a video feed using the video_source and video_output nodes. See Camera Streaming & Multimedia for valid input/output streams, and substitute your desired input and output argument below. For example, you can use video files for the input or output, or use V4L2 cameras instead of MIPI CSI cameras. You can also use RTP/RTSP streams over the network.

# ROS
$ roslaunch ros_deep_learning video_viewer.ros1.launch input:=csi://0 output:=display://0

# ROS2
$ ros2 launch ros_deep_learning video_viewer.ros2.launch input:=csi://0 output:=display://0

imagenet Node

You can launch a classification demo with the following commands - substitute your desired camera or video path to the input argument below (see here for valid input/output streams).

Note that the imagenet node also publishes classification metadata on the imagenet/classification topic in a vision_msgs/Detection2DArray message -- see the Topics & Parameters section below for more info.

# ROS
$ roslaunch ros_deep_learning imagenet.ros1.launch input:=csi://0 output:=display://0

# ROS2
$ ros2 launch ros_deep_learning imagenet.ros2.launch input:=csi://0 output:=display://0

detectnet Node

To launch an object detection demo, substitute your desired camera or video path to the input argument below (see here for valid input/output streams). Note that the detectnet node also publishes the metadata in a vision_msgs/Detection2DArray message -- see the Topics & Parameters section below for more info.

# ROS
$ roslaunch ros_deep_learning detectnet.ros1.launch input:=csi://0 output:=display://0

# ROS2
$ ros2 launch ros_deep_learning detectnet.ros2.launch input:=csi://0 output:=display://0

segnet Node

To launch a semantic segmentation demo, substitute your desired camera or video path to the input argument below (see here for valid input/output streams). Note that the segnet node also publishes raw segmentation results to the segnet/class_mask topic -- see the Topics & Parameters section below for more info.

# ROS
$ roslaunch ros_deep_learning segnet.ros1.launch input:=csi://0 output:=display://0

# ROS2
$ ros2 launch ros_deep_learning segnet.ros2.launch input:=csi://0 output:=display://0

Topics & Parameters

Below are the message topics and parameters that each node implements.

imagenet Node

detectnet Node

segnet Node

video_source Node

video_output Node