ROS package for Coral Edge TPU USB Accelerator

• Ubuntu 16.04 + Kinetic
• Ubuntu 18.04 + Melodic

If you want to run this on Ubuntu 14.04 + Indigo, please see indigo branch.

If you want to run this on PR2, please see pr2 branch.

We need python3.5 or python3.6 to run this package.

For more information, please see here.

For more information, please see here.

For more information, please see here.

For more information, please see here.

For more information, please see here.

echo "deb https://packages.cloud.google.com/apt coral-edgetpu-stable main" | sudo tee /etc/apt/sources.list.d/coral-edgetpu.list
curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add -
sudo apt-get update
# If you do not have USB3, install libedgetpu1-legacy-std
sudo apt-get install libedgetpu1-legacy-max
sudo apt-get install python3-edgetpu

sudo apt-get install python3-pip
wget https://dl.google.com/coral/python/tflite_runtime-1.14.0-cp35-cp35m-linux_x86_64.whl
pip3 install tflite_runtime-1.14.0-cp35-cp35m-linux_x86_64.whl

sudo apt-get install python3-pip
wget https://dl.google.com/coral/python/tflite_runtime-1.14.0-cp36-cp36m-linux_x86_64.whl
pip3 install tflite_runtime-1.14.0-cp36-cp36m-linux_x86_64.whl

For more information, please see here.

sudo apt-get install python3-catkin-pkg-modules python3-rospkg-modules python3-venv python3-empy
sudo apt-get install ros-kinetic-opencv3
sudo apt-get install ros-kinetic-catkin
pip3 install --user opencv-python==4.2.0.32 numpy\<1.19.0
source /opt/ros/kinetic/setup.bash
mkdir -p ~/coral_ws/src
cd ~/coral_ws/src
git clone https://github.com/knorth55/coral_usb_ros.git
wstool init
wstool merge coral_usb_ros/fc.rosinstall
wstool merge coral_usb_ros/fc.rosinstall.kinetic
wstool update
rosdep install --from-paths . --ignore-src -y -r
cd ~/coral_ws
catkin init
catkin config -DPYTHON_EXECUTABLE=/usr/bin/python3 -DPYTHON_INCLUDE_DIR=/usr/include/python3.5m -DPYTHON_LIBRARY=/usr/lib/x86_64-linux-gnu/libpython3.5m.so
catkin build

sudo apt-get install python3-catkin-pkg-modules python3-rospkg-modules python3-venv python3-empy
sudo apt-get install python3-opencv
sudo apt-get install ros-melodic-catkin
source /opt/ros/melodic/setup.bash
mkdir -p ~/coral_ws/src
cd ~/coral_ws/src
git clone https://github.com/knorth55/coral_usb_ros.git
wstool init
wstool merge coral_usb_ros/fc.rosinstall
wstool merge coral_usb_ros/fc.rosinstall.melodic
wstool update
rosdep install --from-paths . --ignore-src -y -r
cd ~/coral_ws
catkin init
catkin config -DPYTHON_EXECUTABLE=/usr/bin/python3 -DPYTHON_INCLUDE_DIR=/usr/include/python3.6m -DPYTHON_LIBRARY=/usr/lib/x86_64-linux-gnu/libpython3.6m.so
catkin build

source ~/coral_ws/devel/setup.bash
roscd coral_usb/scripts
rosrun coral_usb download_models.py

Please see here for more detailed information.

roscore

# source normal workspace, not edge tpu workspace
# /opt/ros/kinetic/setup.bash or /opt/ros/melodic/setup.bash
# source /opt/ros/kinetic/setup.bash
source /opt/ros/melodic/setup.bash
rosrun jsk_perception image_publisher.py _file_name:=$(rospack find jsk_perception)/sample/object_detection_example_1.jpg

# source edge tpu workspace
source /opt/ros/${ROS_DISTRO}/setup.bash # THIS IS VERY IMPORTANT FOR MELODIC to set /opt/ros/${ROS_DISTRO}/lib/python2.7/dist-packages in $PYTHONPATH
source ~/coral_ws/devel/setup.bash       # THIS PUT devel/lib/python3/dist-packages in fornt of /opt/ros/${ROS_DISTRO}/lib/python2.7/dist-package
# object detector
roslaunch coral_usb edgetpu_object_detector.launch INPUT_IMAGE:=/image_publisher/output
# face detector
roslaunch coral_usb edgetpu_face_detector.launch INPUT_IMAGE:=/image_publisher/output
# human pose estimator
roslaunch coral_usb edgetpu_human_pose_estimator.launch INPUT_IMAGE:=/image_publisher/output
# semantic segmenter
roslaunch coral_usb edgetpu_semantic_segmenter.launch INPUT_IMAGE:=/image_publisher/output

To subscribe compressed input image, use IMAGE_TRANSPORT:=compressed

roslaunch edgetpu_object_detector.launch INPUT_IMAGE:=/image_publisher/output IMAGE_TRANSPORT:=compressed

# source normal workspace, not edge tpu workspace
# /opt/ros/kinetic/setup.bash or /opt/ros/melodic/setup.bash
source /opt/ros/kinetic/setup.bash
# object detector
rosrun image_view image_view image:=/edgetpu_object_detector/output/image
# face detector
rosrun image_view image_view image:=/edgetpu_face_detector/output/image
# human pose estimator
rosrun image_view image_view image:=/edgetpu_human_pose_estimator/output/image
# semantic segmenter
rosrun image_view image_view image:=/edgetpu_semantic_segmenter/output/image

To subscribe compressed output image, set ~image_transport param to compressed

rosrun image_view image_view image:=/edgetpu_object_detector/output/image _image_transport:=compressed

• ~input/image (sensor_msgs/Image)

  • Input image

• ~output/rects (jsk_recognition_msgs/RectArray)

  • Rectangles of detected objects

• ~output/class (jsk_recognition_msgs/ClassificationResult)

  • Classification results of detected objects

• ~output/image (sensor_msgs/Image)

  • Visualization of detection results

• ~classifier_name (String, default: rospy.get_name())

  • Classifier name

• ~model_file (String, default: $(rospack find coral_usb)/models/mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite)

  • Model file path

• ~label_file (String, default: $(rospack find coral_usb)/models/coco_labels.txt)

  • Label file path.

• ~enable_visualization (Bool, default: True)

  • Whether enable visualization or not

• ~visualize_duration (Float, default: 0.1)

  • Time duration for visualization

• ~image_transport: (String, default: raw)

  • Set compressed to subscribe compressed image

• ~score_thresh: (Float, default: 0.6)

  • Score threshold for object detection

• ~top_k: (Int, default: 100)

  • Maximum number of detected objects

• ~input/image (sensor_msgs/Image)

  • Input image

• ~output/rects (jsk_recognition_msgs/RectArray)

  • Rectangles of detected faces

• ~output/class (jsk_recognition_msgs/ClassificationResult)

  • Classification results of detected faces

• ~output/image (sensor_msgs/Image)

  • Visualization of detection results

• ~classifier_name (String, default: rospy.get_name())

  • Classifier name

• ~model_file (String, default: $(rospack find coral_usb)/models/mobilenet_ssd_v2_face_quant_postprocess_edgetpu.tflite)

  • Model file path

• ~enable_visualization (Bool, default: True)

  • Whether enable visualization or not

• ~visualize_duration (Float, default: 0.1)

  • Time duration for visualization

• ~image_transport: (String, default: raw)

  • Set compressed to subscribe compressed image

• ~score_thresh: (Float, default: 0.6)

  • Score threshold for face detection

• ~top_k: (Int, default: 100)

  • Maximum number of detected faces

• ~input/image (sensor_msgs/Image)

  • Input image

• ~output/poses (jsk_recognition_msgs/PeoplePoseArray)

  • Estimated human poses

• ~output/rects (jsk_recognition_msgs/RectArray)

  • Rectangles of detected humans

• ~output/class (jsk_recognition_msgs/ClassificationResult)

  • Classification results of detected humans

• ~output/image (sensor_msgs/Image)

  • Visualization of estimation results

• ~classifier_name (String, default: rospy.get_name())

  • Classifier name

• ~model_file (String, default: $(rospack find coral_usb)/python/coral_usb/posenet/models/posenet_mobilenet_v1_075_481_641_quant_decoder_edgetpu.tflite)

  • Model file path

• ~enable_visualization (Bool, default: True)

  • Whether enable visualization or not

• ~visualize_duration (Float, default: 0.1)

  • Time duration for visualization

• ~image_transport: (String, default: raw)

  • Set compressed to subscribe compressed image

• ~score_thresh: (Float, default: 0.2)

  • Score threshold for human pose estimation

• ~joint_score_thresh: (Float, default: 0.2)

  • Score threshold of each joint for human pose estimation

• ~input/image (sensor_msgs/Image)

  • Input image

• ~output/label (sensor_msgs/Image)

  • Estimated label image

• ~output/image (sensor_msgs/Image)

  • Visualization of estimation results

• ~classifier_name (String, default: rospy.get_name())

  • Classifier name

• ~model_file (String, default: $(rospack find coral_usb)/models/deeplabv3_mnv2_pascal_quant_edgetpu.tflite)

  • Model file path

• ~label_file (String, default: None)

  • Label file path. pascal_voc label is used by default.

• ~enable_visualization (Bool, default: True)

  • Whether enable visualization or not

• ~visualize_duration (Float, default: 0.1)

  • Time duration for visualization

• ~image_transport: (String, default: raw)

  • Set compressed to subscribe compressed image

You can see the sample launch edgetpu_node_manager.launch

• ~nodes: (Dict, default: {})

  • Dictionary of node's name and type.
  • type can be as follow:
    • edgetpu_object_detector
    • edgetpu_face_detector
    • edgetpu_human_pose_estimator
    • edgetpu_semantic_segmenter
  • Parameters for each node can be set after name namespace.

• ~default: (String, default: None)

  • Default node name

• ~prefix: (String, default: '')

  • Prefix for each nodes

• ~start: (coral_usb/StartNode)

  • Start node by node name

• ~stop: (coral_usb/StopNode)

  • Stop node