The OpenVINO™ (Open visual inference and neural network optimization) toolkit provides a ROS-adaptered runtime framework of neural network which quickly deploys applications and solutions for vision inference. By leveraging Intel® OpenVINO™ toolkit and corresponding libraries, this runtime framework extends workloads across Intel® hardware (including accelerators) and maximizes performance.

• Enables CNN-based deep learning inference at the edge
• Supports heterogeneous execution across computer vision accelerators—CPU, GPU, Intel® Movidius™ Neural Compute Stick, and FPGA—using a common API
• Speeds up time to market via a library of functions and preoptimized kernels
• Includes optimized calls for OpenCV and OpenVX*

From the view of hirarchical architecture design, the package is divided into different functional components, as shown in below picture.

• Intel® OpenVINO™ toolkit is leveraged to provide deep learning basic implementation for data inference. is free software that helps developers and data scientists speed up computer vision workloads, streamline deep learning inference and deployments, and enable easy, heterogeneous execution across Intel® platforms from edge to cloud. It helps to:
  • Increase deep learning workload performance up to 19x1 with computer vision accelerators from Intel.
  • Unleash convolutional neural network (CNN)-based deep learning inference using a common API.
  • Speed development using optimized OpenCV* and OpenVX* functions.
• ROS2 OpenVINO Runtime Framework is the main body of this repo. it provides key logic implementation for pipeline lifecycle management, resource management and ROS system adapter, which extends Intel OpenVINO toolkit and libraries. Furthermore, this runtime framework provides ways to ease launching, configuration and data analytics and re-use.
• Diversal Input resources are the data resources to be infered and analyzed with the OpenVINO framework.
• ROS interfaces and outputs currently include Topic and service. Natively, RViz output and CV image window output are also supported by refactoring topic message and inferrence results.
• Optimized Models provides by Model Optimizer component of Intel® OpenVINO™ toolkit. Imports trained models from various frameworks (Caffe*, Tensorflow*, MxNet*, ONNX*, Kaldi*) and converts them to a unified intermediate representation file. It also optimizes topologies through node merging, horizontal fusion, eliminating batch normalization, and quantization.It also supports graph freeze and graph summarize along with dynamic input freezing.

From the view of logic implementation, the package introduces the definitions of parameter manager, pipeline and pipeline manager. The below picture depicts how these entities co-work together when the corresponding program is launched.

Once a corresponding program is launched with a specified .yaml config file passed in the .launch.py file or via commandline, parameter manager analyzes the configurations about pipeline and the whole framework, then shares the parsed configuration information with pipeline procedure. A pipeline instance is created by following the configuration info and is added into pipeline manager for lifecycle control and inference action triggering.

The contents in .yaml config file should be well structured and follow the supported rules and entity names. Please see the configuration guidance for how to create or edit the config files.

Pipeline fulfills the whole data handling process: initiliazing Input Component for image data gathering and formating; building up the structured inference network and passing the formatted data through the inference network; transfering the inference results and handling output, etc.

Pipeline manager manages all the created pipelines according to the inference requests or external demands (say, system exception, resource limitation, or end user's operation). Because of co-working with resource management and being aware of the whole framework, it covers the ability of performance optimization by sharing system resource between pipelines and reducing the burden of data copy.

Currently, the package support several kinds of input resources of gaining image data:

Currently, the inference feature list is supported:

• Image topic: /openvino_toolkit/image_raw(sensor_msgs::msg::Image)

• Face Detection: /ros2_openvino_toolkit/face_detection(object_msgs:msg:ObjectsInBoxes)
• Emotion Recognition: /ros2_openvino_toolkit/emotions_recognition(people_msgs:msg:EmotionsStamped)
• Age and Gender Recognition: /ros2_openvino_toolkit/age_genders_Recognition(people_msgs:msg:AgeGenderStamped)
• Head Pose Estimation: /ros2_openvino_toolkit/headposes_estimation(people_msgs:msg:HeadPoseStamped)
• Object Detection: /ros2_openvino_toolkit/detected_objects(object_msgs::msg::ObjectsInBoxes)
• Object Segmentation: /ros2_openvino_toolkit/segmented_obejcts(people_msgs::msg::ObjectsInMasks)
• Person Reidentification: /ros2_openvino_toolkit/reidentified_persons(people_msgs::msg::ReidentificationStamped)
• Rviz Output: /ros2_openvino_toolkit/image_rviz(sensor_msgs::msg::Image)

• Object Detection Service: /detect_object (object_msgs::srv::DetectObject)
• Face Detection Service: /detect_face (object_msgs::srv::DetectObject)
• Age & Gender Detection Service: /detect_age_gender (people_msgs::srv::AgeGender)
• Headpose Detection Service: /detect_head_pose (people_msgs::srv::HeadPose)
• Emotion Detection Service: /detect_emotion (people_msgs::srv::Emotion)

RViz dispaly is also supported by the composited topic of original image frame with inference result. To show in RViz tool, add an image marker with the composited topic: /ros2_openvino_toolkit/image_rviz(sensor_msgs::msg::Image)

OpenCV based image window is natively supported by the package. To enable window, Image Window output should be added into the output choices in .yaml config file. see the config file guidance for checking/adding this feature in your launching.

See below pictures for the demo result snapshots.

• face detection input from image

• object detection input from realsense camera

• object segmentation input from video

• Person Reidentification input from standard camera

NOTE: Intel releases 2 different series of OpenVINO Toolkit, we call them as OpenSource Version and Tarball Version. This guidelie uses OpenSource Version as the installation and launching example. If you want to use Tarball version, please follow the guide for Tarball Version.

1. Intel Distribution of OpenVINO toolkit
   

   • Download OpenVINO toolkit by following the guide
     

   cd ~/Downloads
   wget -c http://registrationcenter-download.intel.com/akdlm/irc_nas/15078/l_openvino_toolkit_p_2018.5.455.tgz

   • Install OpenVINO toolkit by following the guide
     

   cd ~/Downloads
   tar -xvf l_openvino_toolkit_p_2018.5.455.tgz
   cd l_openvino_toolkit_p_2018.5.455
   # root is required instead of sudo
   sudo -E ./install_cv_sdk_dependencies.sh
   sudo ./install_GUI.sh
   # build sample code under OpenVINO toolkit
   source /opt/intel/computer_vision_sdk/bin/setupvars.sh
   cd /opt/intel/computer_vision_sdk/deployment_tools/inference_engine/samples/
   mkdir build
   cd build
   cmake ..
   make

   • Configure the Neural Compute Stick USB Driver

   cd ~/Downloads
   cat <<EOF > 97-usbboot.rules
   SUBSYSTEM=="usb", ATTRS{idProduct}=="2150", ATTRS{idVendor}=="03e7", GROUP="users", MODE="0666", ENV{ID_MM_DEVICE_IGNORE}="1"
   SUBSYSTEM=="usb", ATTRS{idProduct}=="2485", ATTRS{idVendor}=="03e7", GROUP="users", MODE="0666", ENV{ID_MM_DEVICE_IGNORE}="1"
   SUBSYSTEM=="usb", ATTRS{idProduct}=="f63b", ATTRS{idVendor}=="03e7", GROUP="users", MODE="0666", ENV{ID_MM_DEVICE_IGNORE}="1"
   EOF
   sudo cp 97-usbboot.rules /etc/udev/rules.d/
   sudo udevadm control --reload-rules
   sudo udevadm trigger
   sudo ldconfig
   rm 97-usbboot.rules

2. Configure the environment (you can write the configuration to your ~/.basrch file)
   Note: If you used root privileges to install the OpenVINO binary package, it installs the Intel Distribution of OpenVINO toolkit in this directory: /opt/intel/openvino_/

   export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/intel/computer_vision_sdk/deployment_tools/inference_engine/samples/build/intel64/Release/lib
   source /opt/intel/computer_vision_sdk/bin/setupvars.sh

One-step installation scripts are provided for the dependencies' installation. Please see the guide for details.

• Preparation
  • download and convert a trained model to produce an optimized Intermediate Representation (IR) of the model

     #object segmentation model
     cd /opt/openvino_toolkit/dldt/model-optimizer/install_prerequisites
     sudo ./install_prerequisites.sh
     mkdir -p ~/Downloads/models
     cd ~/Downloads/models
     wget http://download.tensorflow.org/models/object_detection/mask_rcnn_inception_v2_coco_2018_01_28.tar.gz
     tar -zxvf mask_rcnn_inception_v2_coco_2018_01_28.tar.gz
     cd mask_rcnn_inception_v2_coco_2018_01_28
     python3 /opt/openvino_toolkit/dldt/model-optimizer/mo_tf.py --input_model frozen_inference_graph.pb --tensorflow_use_custom_operations_config /opt/openvino_toolkit/dldt/model-optimizer/extensions/front/tf/mask_rcnn_support.json --tensorflow_object_detection_api_pipeline_config pipeline.config --reverse_input_channels --output_dir ./output/
     sudo mkdir -p /opt/models
     sudo ln -s ~/Downloads/models/mask_rcnn_inception_v2_coco_2018_01_28 /opt/models/
     #object detection model
     cd /opt/openvino_toolkit/open_model_zoo/model_downloader
     python3 downloader.py --name ssd300
     sudo python3 /opt/openvino_toolkit/dldt/model-optimizer/mo.py --input_model /opt/openvino_toolkit/open_model_zoo/model_downloader/object_detection/common/ssd/300/caffe/ssd300.caffemodel --output_dir /opt/openvino_toolkit/open_model_zoo/model_downloader/object_detection/common/ssd/300/caffe/output/

  • download the optimized Intermediate Representation (IR) of model (excute once)
    

     cd /opt/openvino_toolkit/open_model_zoo/model_downloader
     python3 downloader.py --name face-detection-adas-0001
     python3 downloader.py --name age-gender-recognition-retail-0013
     python3 downloader.py --name emotions-recognition-retail-0003
     python3 downloader.py --name head-pose-estimation-adas-0001
     python3 downloader.py --name person-detection-retail-0013
     python3 downloader.py --name person-reidentification-retail-0076

  • copy label files (excute once)
    

     sudo cp /opt/openvino_toolkit/ros2_openvino_toolkit/data/labels/emotions-recognition/FP32/emotions-recognition-retail-0003.labels /opt/openvino_toolkit/open_model_zoo/model_downloader/Retail/object_attributes/emotions_recognition/0003/dldt
     sudo cp /opt/openvino_toolkit/ros2_openvino_toolkit/data/labels/face_detection/face-detection-adas-0001.labels /opt/openvino_toolkit/open_model_zoo/model_downloader/Transportation/object_detection/face/pruned_mobilenet_reduced_ssd_shared_weights/dldt
     sudo cp /opt/openvino_toolkit/ros2_openvino_toolkit/data/labels/object_segmentation/frozen_inference_graph.labels /opt/models/mask_rcnn_inception_v2_coco_2018_01_28/output
     sudo cp /opt/openvino_toolkit/ros2_openvino_toolkit/data/labels/object_detection/ssd300.labels /opt/openvino_toolkit/open_model_zoo/model_downloader/object_detection/common/ssd/300/caffe/output

  • set ENV LD_LIBRARY_PATH
    

     export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/openvino_toolkit/dldt/inference-engine/bin/intel64/Release/lib

• run face detection sample code input from StandardCamera.

   ros2 launch dynamic_vino_sample pipeline_people_oss.launch.py

• run face detection sample code input from Image.

   ros2 launch dynamic_vino_sample pipeline_image_oss.launch.py

• run object detection sample code input from RealSenseCamera.

   ros2 launch dynamic_vino_sample pipeline_object_oss.launch.py

• run object detection sample code input from RealSenseCameraTopic.

   ros2 launch dynamic_vino_sample pipeline_object_oss_topic.launch.py

• run object segmentation sample code input from RealSenseCameraTopic.

   ros2 launch dynamic_vino_sample pipeline_segmentation.launch.py

• run object segmentation sample code input from Video.

   ros2 launch dynamic_vino_sample pipeline_video.launch.py

• run person reidentification sample code input from StandardCamera.

   ros2 launch dynamic_vino_sample pipeline_reidentification_oss.launch.py

• run object detection service sample code input from Image
  Run image processing service:

   ros2 launch dynamic_vino_sample image_object_server_oss.launch.py

  Run example application with an absolute path of an image on another console:

   ros2 run dynamic_vino_sample image_object_client ~/Pictures/car.png

• run face detection service sample code input from Image
  Run image processing service:

   ros2 launch dynamic_vino_sample image_people_server_oss.launch.py

  Run example application with an absolute path of an image on another console:

   ros2 run dynamic_vino_sample image_people_client ~/Pictures/face.png

• Support result filtering for inference process, so that the inference results can be filtered to different subsidiary inference. For example, given an image, firstly we do Object Detection on it, secondly we pass cars to vehicle brand recognition and pass license plate to license number recognition.
• Design resource manager to better use such resources as models, engines, and other external plugins.
• Develop GUI based configuration and management tools (and monitoring and diagnose tools), in order to provide easy entry for end users to simplify their operation.

• ROS2 OpenVINO discription writen in Chinese: https://mp.weixin.qq.com/s/BgG3RGauv5pmHzV_hkVAdw