ros2_coco_detector

Integrate PyTorch Torchvision MobileNet for Microsoft COCO object detection into the ROS2 environment

YouTube Demonstration

Summary

This package performs object detection in the ROS2 environment. There are certainly more sophisticated object detection frameworks out there, eg. NVIDIA Isaac.

The chief virtue of this package is the simplicity of the codebase and use of standardised ROS2 messages, with the goal of being simple to understand and to use.

Packages

coco_detector: package containing coco_detector_node for listening on ROS2 topic /image and publishing ROS2 Detection2DArray message on topic /detected_objects. Also (by default) publishes Image (with labels and bounding boxes) message on topic /annotated_image. The object detection is performed by PyTorch using MobileNet.

Tested Hardware

Dell Precision Tower 2210, NVIDIA RTX2070 (GPU is optional)

Tested Software

Ubuntu 22.04, ROS2 Humble (RoboStack), PyTorch 2.1.2, CUDA 12.2 (CUDA is only needed if you require GPU)

Installation - historical reference only

NOTE - outdated - not relevant - for historial reference only.

Follow the RoboStack installation instructions to install ROS2

(Ensure you have also followed the step Installation tools for local development in the above instructions)

Follow the PyTorch installation instructions to install PyTorch (selecting the conda option).

mamba activate ros2  # (use the name here you decided to call this conda environment)
mamba install ros-humble-image-tools
mamba install ros-humble-vision-msgs
cd ~
mkdir -p ros2_ws/src
cd ros2_ws
git -C src clone https://github.com/jfrancis71/ros2_coco_detector.git
colcon build --symlink-install

You may receive a warning on the colcon build step: “SetuptoolsDeprecationWarning: setup.py install is deprecated”, this can be ignored.

The above steps assume a RoboStack mamba/conda ROS2 install. If using other installation process, replace the RoboStack image-tools and vision-msgs packages install steps with whichever command is appropriate for your environment. The image-tools package is not required for coco_detector, it is just used in the steps below for convenient demonstration. However vision-msgs is required (this is where the ROS2 DetectionArray2D message is defined)

Activate Environment

mamba activate ros2 # (use the name here you decided to call this conda environment)
cd ~/ros2_ws
source ./install/setup.bash

Verify Install

Launch a camera stream:

ros2 run image_tools cam2image

On another terminal enter:

ros2 run coco_detector coco_detector_node

There will be a short delay the first time the node is run for PyTorch TorchVision to download the neural network. You should see a downloading progress bar. This network is then cached for subsequent runs.

On another terminal to view the detection messages:

ros2 topic echo /detected_objects

To view the image stream annotated with the labels and bounding boxes:

ros2 run image_tools showimage --ros-args -r /image:=/annotated_image

Example Use:

ros2 run coco_detector coco_detector_node --ros-args -p publish_annotated_image:=False -p device:=cuda -p detection_threshold:=0.7

This will run the coco detector without publishing the annotated image (it is True by default) using the default CUDA device (device=cpu by default). It sets the detection_threshold to 0.7 (it is 0.9 by default). The detection_threshold should be between 0.0 and 1.0; the higher this number the more detections will be rejected. If you have too many false detections try increasing this number. Thus only Detection2DArray messages are published on topic /detected_objects.

Suggested Setup For Mobile Robotics

These suggestions are for a Raspberry Pi 3 Model B+ running ROS2.

As of 16/02/2024, the PyTorch Conda install does not appear to be working for Raspberry Pi 3 Model B+. There may be other installation options, but I have not explored that.

As an alternative if you have a ROS2 workstation connected to the same network, I suggest publishing the compressed image on the Raspberry Pi and running the COCO detector on the workstation.

The below setup involves the ROS2 compression transport on both the Raspberry Pi and workstation. If using RoboStack ROS2 Humble you can install on each with:

```mamba install ros-humble-compressed-image-transport

Raspberry Pi (run each command in seperate terminals):


```ros2 run image_tools cam2image --ros-args -r /image:=/charlie/image

```ros2 run image_transport republish raw compressed –ros-args -r in:=/charlie/image -r out/compressed:=/charlie/compressed

Workstation (run each command in seperate terminals):


```ros2 run image_transport republish compressed raw --ros-args -r /in/compressed:=/charlie/compressed -r /out:=/server/image

ros2 run coco_detector coco_detector_node --ros-args -r /image:=/server/image

I have relabelled topic names for clarity and keeping the image topics on the different machines seperate. Compression is not necessary, but I have poor performance on my network without compression.

Note you could use launch files (for convenience) to run the above nodes. I do not cover that here as it will be specific to your setup.

Notes

The ROS2 documentation suggests that the ObjectHypotheses.class_id should be an identifier that the client should then look up in a database. This seems more complex than I have a need for. So this implementation just places the class label here directly, eg. class_id = “dog”. See the ROS2 Vision Msgs Github link in the external links section below for more details.

External Links

COCO Dataset Homepage

Microsoft COCO: Common Objects in Context

PyTorch MobileNet

MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

ROS2 Vision Msgs Github repo