airo-camera-toolkit

airo-camera-toolkit

This package contains code for working with RGB(D) cameras, images and point clouds.

Overview of the functionality and the structure:

airo-camera-toolkit/
├── calibration/                # hand-eye extrinsics calibration
├── cameras/                    # actual camera drivers
├── image_transformations/      # reversible geometric 2D transforms
├── pinhole_operations/         # 2D-3D operations
├── point_clouds/               # conversions and operations
├── utils/                      # a.o. annotation tool and converter
├── interfaces.py
└── cli.py

Installation

The airo_camera_toolkit package can be installed with pip by running (from this directory):

pip install .

This will already allow you to use the hardare-independent functionality of this package, e.g. image conversion and projection. Depending on the hardware you are using, you might need to complete additional installation. Instructions can be found in the following files:

• ZED Installation
• RealSense Installation

Additionally, to ensure you have airo-robots installed for the hand-eye calibration, install the extra dependencies:

pip install .[hand-eye-calibration]

Getting started with cameras

Camera can be accessed by instantiating the corresponding class:, e.g. for a ZED camera:

from airo_camera_toolkit.cameras.zed2i import Zed2i
from airo_camera_toolkit.utils import ImageConverter
import cv2

camera = Zed2i(Zed2i.RESOLUTION_720, fps=30)

while True:
    image_rgb_float = camera.get_rgb_image()
    image_bgr = ImageConverter.from_numpy_format(image_rgb_float).image_in_opencv_format
    cv2.imshow("Image", image_bgr)
    key = cv2.waitKey(10)
    if key == ord('q'):
        break

Hand-eye calibration

Find the pose of a camera relative to a robot. In the calibration folder, run this to get started:

airo-camera-toolkit hand-eye-calibration --help

See calibration/README.md for more details.

Utils

Image format conversion

Camera by default return images as numpy 32-bit float RGB images with values between 0 to 1 through get_rgb_image(). This is most convenient for subsequent processing, e.g. with neural networks. For higher performance, 8-bit unsigned integer RGB images are also accessible through get_rgb_image_as_int().

However, when using OpenCV, you will need conversion to BGR format. For this you can use the ImageConverter class:

from airo_camera_toolkit.utils import ImageConverter

image_rgb_int = camera.get_rgb_image_as_int()
image_bgr = ImageConverter.from_numpy_int_format(image_rgb_int).image_in_opencv_format

Annotation tool

See annotation_tool.md for usage instructions.

Pinhole Operations

2D - 3D geometric operations using the pinhole model. See readme for more information.

Image Transforms

See the README in the image_transforms folder for more details.

Real-time visualisation

For realtime visualisation of robotics data we strongly encourage using rerun.io instead of manually hacking something together with opencv/pyqt/... No wrappers are needed here, just pip install the SDK. An example notebook to get to know this tool and its potential can be found here. See this README for more details.

Point clouds

See the tutorial notebook here for an introduction.

Multiprocessing

Camera processing can be computationally expensive. If this is a problem for your application, see multiprocess/README.md.

References

For more background on cameras, in particular on the meaning of intrinsics, extrinics, distortion coefficients, pinhole (and other) camera models, see:

• Szeliski - Computer vision: Algorithms and Applications, available here
• https://web.eecs.umich.edu/~justincj/teaching/eecs442/WI2021/schedule.html
• https://learnopencv.com/geometry-of-image-formation/ (extrinsics & intrinsics)
• http://www.cs.cmu.edu/~16385/s17/Slides/11.1_Camera_matrix.pdf (idem)