Perception for marine surface vehicles

• CPU: i7 9th gen or later recommended
• RAM: 8GB+
• Video card: 5GB+ Video RAM, CUDA 10 compatible
• Storage: ~20GB for docker images and ML models
• O/S: Linux, Docker 18.03 or later, CUDA drivers
• ROS Melodic host/image with a TF tree defined, including a fixed frame (eg, odom)
• IMU publisher, ENU, type: sensor_msgs::Imu
• RADAR image, type sensor_msgs::CompressedImage
• RADAR pointcloud publisher, type sensor_msgs::PointCloud2
• Camera image publisher, color, type sensor_msgs::CompressedImage

asv_perception contains ROS nodes specialized for image processing, pointcloud generation, obstacle creation, and obstacle tracking. The expected inputs are described in the System Requirements and the fused/tracked obstacles are published in the /tracking/fusion/obstacles ROS topic. See packages/asv_perception_common/msg/Obstacle.msg for a list of the fields which are included.

The system can currently accommodate visible light cameras and pointcloud-generating sensors (eg RADAR, LIDAR), and multiple sensors per type. Additionally, an example tool is provided for publishing the obstacle information in JSON format to an external source/system over IP. See the Individual package description section below additional details.

The default configuration is for a monocular camera, LIDAR, and two RADAR sensors, but can be adjusted to accommodate alternative sensor configurations.

• git clone --recursive https://github.com/uml-marine-robotics/asv_perception.git
• cd asv_perception
• Adjust launch files for your sensor configuration
• Put classifier data in data/classification/
  • To use Darknet MS-COCO pretrained weights:
    • Download appropriate cfg and weights from https://github.com/AlexeyAB/darknet/wiki/YOLOv4-model-zoo to model.cfg and model.weights
    • Download https://github.com/AlexeyAB/darknet/blob/master/cfg/coco.names to model.names
    • Download https://github.com/AlexeyAB/darknet/blob/master/cfg/coco.data to model.data
      • In this file, change names = ... to names = /data/classification/model.names
• (Optional) Put WaSR segmentation weights in data/segmentation/
  • Only required if using segmentation node with USE_SEGMENTATION environment variable
  • Pretrained segmentation weights can be acquired from https://github.com/bborja/wasr_network
• ./build.sh

• Calibration tool:
  • Start asv_perception image with the VNC_PORT environment variable specified (e.g. VNC_PORT=5900). In other words, execute run.sh by '. run.sh'. Make sure that the
    docker container is running successfully.
  • Start robot sensor ROS drivers, or start a .bag file (rosbag play )
  • Connect to the asv_perception GUI with a vnc client, at <docker_ip_address>::<VNC_PORT>. A desktop window should appear within VNC. If you are on Ubuntu Linux, 'vncviewer' should be available to you. If you are running the VNC client from the same machine as the docker container, then you need not give IP address. Just specifying port number is sufficient.
  • In the VNC desktop, click Run Calibration. A rqt_image_view window and a Calibration window should appear
  • Change rqt_image_view to cameraN/homography_vis/image, where cameraN is the camera you wish to calibrate. You should see the camera image along with the RADAR image overlay
  • In the Calibration window, set the topic to cameraN/homography, where cameraN is the camera you wish to calibrate.
  • Adjust the radar image using the tuner until sufficient alignment is achieved.
  • When complete, save the calibration, then close all windows and exit VNC.

./run.sh to start the asv_perception nodes, then run a .bag file or connect to asv sensors. Use rviz and/or rqt_image_view to display output. See ./run.sh for command line options.

• docker: Build files and data for docker images
• packages: ROS packages
• data: Configuration data and deep learning models

Each node may have multiple inputs, outputs, and parameters to control their behavior. Below is a high level summary of each node's purpose and main ROS nodes. See the referenced file headers for details on input/output topics and parameters.

• asv_perception_common: Common messages and code shared by other packages
• asv_perception_homography: Integrates IMU and camera-radar homography data. Publishes homography matrix for camera to radar and supports the calibration tool
  • Node: nodes/homography.py
  • Node: nodes/visualization.py
  • Calibration tool: See Setup/calibration section above
• asv_perception_classification: Performs object detection in camera images and reports the ROI
  • Node: src/darknet_node.py
• asv_perception_segmentation: Semantic separation of camera image into water, sky, and obstacle components, publishing an image for each component.
  • Node: src/segmentation_node.py
• asv_perception_obstacleid: Classified obstacle projection, unclassified obstacle pixel projection and clustering, custom Point Cloud filtering and concatenation.
  • Node: include/ObstacleExtractionNodelet.h, euclidean clustering and constructs asv_perception_common/Obstacle messages
  • Node: include/ObstacleProjectionNodelet.h, classified bounding box adjustment, Obstacle message creation, unclassified pixel projection
  • Node: include/PointCloudConcatNodelet.h, Concatenates multiple partial pointcloud segments received over time into a single pointcloud
  • Node: include/PointCloudFilterNodelet.h, Custom pointcloud filtering, including min-distance, outlier removal, and cluster area inclusion/exclusion
• asv_perception_tracking: Sensor-level obstacle tracking and fusion of multi-sensor obstacles. RViz visualization of obstacles. Reporting of obstacle data to external systems
  • Node: nodes/fusion.py: Fusion of tracked obstacle data from N sensors into singular output
  • Node: nodes/nmea_reporting.py: Example of obstacle reporting in NMEA format
  • Node: nodes/socket_reporting.py: Example of obstacle reporting in JSON format over UDP
  • Tool: nodes/socket_reporting_listener.py: Example of listener application which receives JSON data over UDP
  • Node: nodes/tracking.py: Node which performs sensor-level obstacle tracking
  • Node: nodes/visualization.py: Obstacle visualization in RViz