coral_usb_ros

ROS package for Coral Edge TPU USB Accelerator

Environment

• Ubuntu 16.04 + Kinetic
• Ubuntu 18.04 + Melodic
• Ubuntu 20.04 + Noetic

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.

Notice

We need python3.5 and above to run this package.

ROS node list

Object detector: edgetpu_object_detector.py

For more information, please see here.

Face detector: edgetpu_face_detector.py

For more information, please see here.

Human pose estimator: edgetpu_human_pose_estimator.py

For more information, please see here.

Semantic segmenter: edgetpu_semantic_segmenter.py

For more information, please see here.

Panorama object detector: edgetpu_panorama_object_detector.py

For more information, please see here.

Panorama face detector: edgetpu_panorama_face_detector.py

For more information, please see here.

Panorama human pose estimator: edgetpu_panorama_human_pose_estimator.py

For more information, please see here.

Panorama semantic segmenter: edgetpu_panorama_semantic_segmenter.py

For more information, please see here.

Node manager: edgetpu_node_manager.py

EdgeTPU node manager to start, stop and switch multiple nodes.

This node manager is useful if you have limited TPU resource.

List available and running nodes

$ rosservice call /edgetpu_node_manager/list
running_node_name: "edgetpu_object_detector"
node_names:
  - edgetpu_object_detector
  - edgetpu_panorama_object_detector

Start or switch node

$ rosservice call /edgetpu_node_manager/start "name: 'edgetpu_object_detector'"
success: True
last_node_name: "edgetpu_panorama_object_detector"

Stop node

$ rosservice call /edgetpu_node_manager/stop "{}"
success: True
last_node_name: "edgetpu_object_detector"

For more information, please see here.

Setup

Edge TPU dependencies installation

Install the Edge TPU runtime

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 # Choose <YES> when asked
sudo apt-get install python3-edgetpu

Install just the TensorFlow Lite interpreter (Kinetic)

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

Install just the TensorFlow Lite interpreter (Melodic)

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

Install just the TensorFlow Lite interpreter (Noetic)

sudo apt-get install python3-tflite-runtime

For more information, please see here.

Workspace build

Workspace build (Kinetic)

Kinetic workspace requires cv-bridge build with python3.5.

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/jsk-ros-pkg/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

Workspace build (Melodic + cv-bridge-python3)

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/jsk-ros-pkg/coral_usb_ros.git
rosdep install --from-paths . --ignore-src -y -r
cd ~/coral_ws
catkin init
catkin build

Workspace build (Noetic)

sudo apt-get install ros-noetic-catkin
source /opt/ros/noetic/setup.bash
mkdir -p ~/coral_ws/src
cd ~/coral_ws/src
git clone https://github.com/jsk-ros-pkg/coral_usb_ros.git
rosdep install --from-paths . --ignore-src -y -r
cd ~/coral_ws
catkin init
catkin build

Model download

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

Model training with your dataset

Please see here for more detailed information.

Add Device Access Permission

You need to your accout to plugdev group. To enable this feature, you need to re-loggin or run exec su -l $(whoami).

sudo adduser $(whoami) plugdev

Demo

Please choose one of these demo.

• USB or Laptop camera demo
• No camera demo
• Insta 360 air demo

USB or Laptop camera demo

Run all following commands in different terminals in parallel.

Run roscore

roscore

Run usb_cam for normal image

# source normal workspace, not edge tpu workspace
source /opt/ros/${ROS_DISTRO}/setup.bash
rosrun usb_cam usb_cam_node

Run Edge TPU launch

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

To subscribe compressed input image, use IMAGE_TRANSPORT:=compressed

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

Run image_view

# source normal workspace, not edge tpu workspace
source /opt/ros/${ROS_DISTRO}/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
# panorama object detector
rosrun image_view image_view image:=/edgetpu_panorama_object_detector/output/image
# panorama face detector
rosrun image_view image_view image:=/edgetpu_panorama_face_detector/output/image
# panorama human pose estimator
rosrun image_view image_view image:=/edgetpu_panorama_human_pose_estimator/output/image
# panorama semantic segmenter
rosrun image_view image_view image:=/edgetpu_panorama_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

No camera demo

Run all following commands in different terminals in parallel.

Run roscore

roscore

Run image_publisher for virtual camera

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

Run Edge TPU launch

# source edge tpu workspace
# THIS IS VERY IMPORTANT FOR MELODIC to set /opt/ros/${ROS_DISTRO}/lib/python2.7/dist-packages in $PYTHONPATH
source /opt/ros/${ROS_DISTRO}/setup.bash
# THIS PUT devel/lib/python3/dist-packages in fornt of /opt/ros/${ROS_DISTRO}/lib/python2.7/dist-package
source ~/coral_ws/devel/setup.bash
# 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

Run image_view

# source normal workspace, not edge tpu workspace
source /opt/ros/${ROS_DISTRO}/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
# panorama object detector
rosrun image_view image_view image:=/edgetpu_panorama_object_detector/output/image
# panorama face detector
rosrun image_view image_view image:=/edgetpu_panorama_face_detector/output/image
# panorama human pose estimator
rosrun image_view image_view image:=/edgetpu_panorama_human_pose_estimator/output/image
# panorama semantic segmenter
rosrun image_view image_view image:=/edgetpu_panorama_semantic_segmenter/output/image

Insta360 air camera demo

Run all following commands in different terminals in parallel.

Run roscore

roscore

Run insta360 air for panorama image

# source normal workspace, not edge tpu workspace
source /opt/ros/${ROS_DISTRO}/setup.bash
roslaunch jsk_perception sample_insta360_air.launch gui:=false

Run Edge TPU launch

For insta360 air panorama image

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

Run image_view

# source normal workspace, not edge tpu workspace
source /opt/ros/${ROS_DISTRO}/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
# panorama object detector
rosrun image_view image_view image:=/edgetpu_panorama_object_detector/output/image
# panorama face detector
rosrun image_view image_view image:=/edgetpu_panorama_face_detector/output/image
# panorama human pose estimator
rosrun image_view image_view image:=/edgetpu_panorama_human_pose_estimator/output/image
# panorama semantic segmenter
rosrun image_view image_view image:=/edgetpu_panorama_semantic_segmenter/output/image

ROS node information

Object detector: edgetpu_object_detector.py

Subscribing Topic

• ~input/image (sensor_msgs/Image)

  • Input image

Publishing Topic

• ~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

Parameters

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

  • Classifier name

• ~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

• ~always_publish (Bool, default: True)

  • Set false to publish when an object is detected.

Dynamic parameters

• ~score_thresh: (Float, default: 0.6)

  • Score threshold for object detection

• ~top_k: (Int, default: 100)

  • Maximum number of detected objects

• ~model_file (String, default: package://coral_usb/models/mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite)

  • Model file path

• ~label_file (String, default: package://coral_usb/models/coco_labels.txt)

  • Label file path.

Face detector: edgetpu_face_detector.py

Subscribing Topic

• ~input/image (sensor_msgs/Image)

  • Input image

Publishing Topic

• ~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

Parameters

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

  • Classifier name

• ~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

• ~always_publish (Bool, default: True)

  • Set false to publish when a face is detected.

Dynamic parameters

• ~score_thresh: (Float, default: 0.6)

  • Score threshold for face detection

• ~top_k: (Int, default: 100)

  • Maximum number of detected faces

• ~model_file (String, default: package://coral_usb/models/mobilenet_ssd_v2_face_quant_postprocess_edgetpu.tflite)

  • Model file path

Human pose estimator: edgetpu_human_pose_estimator.py

Subscribing Topic

• ~input/image (sensor_msgs/Image)

  • Input image

Publishing Topic

• ~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

Parameters

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

  • Classifier name

• ~model_file (String, default: package://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

• ~always_publish (Bool, default: True)

  • Set false to publish when a human pose is detected.

Dynamic parameters

• ~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

Semantic segmenter: edgetpu_semantic_segmenter.py

Subscribing Topic

• ~input/image (sensor_msgs/Image)

  • Input image

Publishing Topic

• ~output/label (sensor_msgs/Image)

  • Estimated label image

• ~output/image (sensor_msgs/Image)

  • Visualization of estimation results

Parameters

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

  • Classifier name

• ~model_file (String, default: package://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

Panorama object detector: edgetpu_panorama_object_detector.py

Subscribing Topic

• ~input/image (sensor_msgs/Image)

  • Input image

Publishing Topic

• ~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

Parameters

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

  • Classifier name

• ~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

• ~always_publish (Bool, default: True)

  • Set false to publish when an object is detected.

Dynamic parameters

• ~score_thresh: (Float, default: 0.6)

  • Score threshold for object detection

• ~top_k: (Int, default: 100)

  • Maximum number of detected objects

• ~model_file (String, default: package://coral_usb/models/mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite)

  • Model file path

• ~label_file (String, default: package://coral_usb/models/coco_labels.txt)

  • Label file path.

• ~n_split (Int, default: 3)

  • Number of splitting images from one large panorama image.

• ~overlap (Bool, default: True)

  • Recognize with overlapping splitted images.

• ~nms (Bool, default: True)

  • Use non-maximum suppression or not for overlap detection.

• ~nms_thresh (Double, default: 0.3)

  • Non-maximum suppression threshold

Panorama face detector: edgetpu_panorama_face_detector.py

Subscribing Topic

• ~input/image (sensor_msgs/Image)

  • Input image

Publishing Topic

• ~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

Parameters

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

  • Classifier name

• ~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

• ~always_publish (Bool, default: True)

  • Set false to publish when a face is detected.

Dynamic parameters

• ~score_thresh: (Float, default: 0.6)

  • Score threshold for face detection

• ~top_k: (Int, default: 100)

  • Maximum number of detected faces

• ~model_file (String, default: package://coral_usb/models/mobilenet_ssd_v2_face_quant_postprocess_edgetpu.tflite)

  • Model file path

• ~n_split (Int, default: 3)

  • Number of splitting images from one large panorama image.

• ~overlap (Bool, default: True)

  • Recognize with overlapping splitted images.

• ~nms (Bool, default: True)

  • Use non-maximum suppression or not for overlap detection.

• ~nms_thresh (Double, default: 0.3)

  • Non-maximum suppression threshold

Panorama human pose estimator: edgetpu_panorama_human_pose_estimator.py

Subscribing Topic

• ~input/image (sensor_msgs/Image)

  • Input image

Publishing Topic

• ~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

Parameters

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

  • Classifier name

• ~model_file (String, default: package://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

• ~always_publish (Bool, default: True)

  • Set false to publish when a human pose is detected.

Dynamic parameters

• ~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

• ~n_split (Int, default: 3)

  • Number of splitting images from one large panorama image.

• ~overlap (Bool, default: True)

  • Recognize with overlapping splitted images.

Panorama semantic segmenter: edgetpu_panorama_semantic_segmenter.py

Subscribing Topic

• ~input/image (sensor_msgs/Image)

  • Input image

Publishing Topic

• ~output/label (sensor_msgs/Image)

  • Estimated label image

• ~output/image (sensor_msgs/Image)

  • Visualization of estimation results

Parameters

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

  • Classifier name

• ~model_file (String, default: package://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

Dynamic parameters

• ~n_split (Int, default: 3)

  • Number of splitting images from one large panorama image.

Node manager: edgetpu_node_manager.py

You can see the sample launch edgetpu_node_manager.launch

Parameters

• ~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
    • edgetpu_panorama_object_detector
    • edgetpu_panorama_face_detector
    • edgetpu_panorama_human_pose_estimator
    • edgetpu_panorama_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

Service

• ~start: (coral_usb/StartNode)

  • Start node by node name

• ~stop: (coral_usb/StopNode)

  • Stop node