Skip to content

Latest commit

 

History

History
160 lines (120 loc) · 6.59 KB

README.md

File metadata and controls

160 lines (120 loc) · 6.59 KB
Raw

DeepMind Robotics Vision

This package provides functions, methods and ROS nodes for colour-based blob detection and 2D-to-3D (from pixel to 3D Cartesian coordinates) triangulation.

Rationale

The goal of this package is to:

  1. detect uniquely coloured objects;
  2. evaluate their image (pixel space) barycenter;
  3. estimate their 3D Cartesian barycenter from multiple images via triangulation.

To run the triangulation algorithm, you need to run the blob detection algorithm on at least two cameras. In ROS, this translates to running two blob detection ROS nodes on two cameras and then the triangulation ROS node on the published pixel barycenters.

Run the ROS nodes

To run the a blob detector ROS node, simply

python launch_blob_detector.py --camera=<camera name>

To run the triangulation ROS node

python launch_blob_triangulation.py

In order to set up the ROS node for your needs, you need to edit the following configuration files:

Name Description
config_blob_detector.py Configure the blob detector ROS node.
config_blob_triangulation.py Configure the blob-based triangulation ROS node.
robot_config.py Robot-specific configuration file.

Configuration files

Each configuration file exposes a getter function that returns a data class containing all the parameters needed to run the algorithm/ROS nodes.

config_blob_detector.py

This modules returns an instance of BlobDetectorConfig that defines the following parameters:

  • node_name: the name of the ROS node.
  • rate: the desired frame rate.
  • input_queue_size: the input data queue size (see ROS documentation).
  • output_queue_size: the output data queue size (see ROS documentation).
  • camera_namespaces: a camera name to ROS namespace mapping.
  • masks: (u, v) coordinates defining closed regions of interest in the image where the blob detector will not look for blobs. Masks are defined as a list of list of points, as required by the pts argument of OpenCV's fillPoly function.
  • scale: image scaling factor to increase speed and frame rate.
  • min_area: minimum size in pixels above which a blob is deemed valid.

To get started, the two parameters to change are camera_namespaces and masks. They both need to have the same key, which corresponds to a camera of your choice, which you will need to pass to the --camera parameter of launch_blob_detector.py.

The camera_namespaces is, as the name and description suggests, the base name of a camera ROS topic. The blob detector will automatically append to it /image_raw and the resulting ROS topic name will be used to collect images and run the blob detector.

The masks must define a mask for each camera. It can be trivially set to either an empty list or None, unless you have a camera configuration where you know a priori where to not look for an object in the image (e.g. in settings where the camera is fixed and the robot workspace is well defined).

robot_config.py

Before jumping into the triangulation configuration file, we need to set up the configuration file for the robot. In fact, the triangulation algorithm depends on some robot parameters and this file collects them.

This modules returns an instance of RobotConfig that defines the following parameters:

  • name: unique robot name.
  • cameras: collection of cameras.
  • basket_center: center of playground relative to the robot base frame in the xy plane.
  • basket_height: displacement of the playground from the robot base frame.
  • base_frame_name: the name (or id) that identifies the robot ROS base frame.

As you can see from the provided config file, the class RobotType can provide multiple RobotConfig, each with a unique name, e.g. STANDARD_SAWYER. When editing or adding a RobotConfig, it is important to note that the camera names must be the same than the ones used in config_blob_detector.py's camera_namespaces. Also note that only the camera extrinsic parameters are required as the triangulation algorithm will collect the intrinsics from the camera ROS topic automatically.

config_blob_triangulation.py

This modules returns an instance of BlobTriangulationConfig that defines the following parameters:

  • node_name: the name of the ROS node.
  • rate: the desired frame rate.
  • input_queue_size: the input data queue size (see ROS documentation).
  • output_queue_size: the output data queue size (see ROS documentation).
  • fuse_tolerance: time in seconds after which data is considered outdated and not used for triangulation.
  • extrinsics: extrinsic camera parameters.
  • limits: Cartesian limits over which points are considered as outliers.
  • base_frame: the name of the robot base frame.
  • deadzones: additional Cartesian limits excluding volumes of the robot operative space where points are discarded.

If you configured the robot_config.py, most of the information can be collected from there and possibly elaborated to populate extrinsics, base_frame, limits, and deadzones. While the first two are quite straightforward, the latter two requires a bit more consideration. Both are used by the triangulation algorithm to validate and possibly reject triangulated points. However, the two parameters carry different semantics. limits is used to describe the robot's workspace volume, while deadzones are forbidden zones. Points are considered valid when inside the limits and outside deadzones.

Launch files

This section describes the launch files and their input arguments for both blob detector and triangulator ROS nodes.

launch_blob_detector.py

Note: the config file is automatically loaded and used.

The arguments are:

  • camera: the camera to use. Must be one of the keys in cameras in the configuration file.
  • props: defaults to RED, GREEN and BLUE objects defined in blob_tracker_object_defs.py. The names of the props to track.
  • visualize: defaults to False. Whether to publish helper images of the detected blobs or not.
  • toolkit: defaults to False. Whether to display a YUV GUI toolkit to find good YUV parameters to detect blobs. Sets visualize = True.

launch_blob_triangulation.py

Note: the config file is automatically loaded and used.

The arguments are:

  • robot: defaults to STANDARD_SAWYER. The name of the robot. Must be one of the enums in the robot configuration file.
  • props: defaults to RED, GREEN and BLUE objects defined in blob_tracker_object_defs.py. The names of the props to track.