This project has software for capturing point clouds using Microsoft Azure Kinect cameras, k4a for short. The software turns k4a depth- and colorframes into a cwipc pointcloud.
It is usually built or installed as part of the cwipc suite, https://github.com/cwi-dis/cwipc.
When installed it enables the --kinect option to tools like cwipc_view, cwipc_grab, cwipc_calibrate and others. This will useKinect cameras as a source of pointclouds.
This module requires the following software to be installed:
- Azure Kinect SDK
- Azure Kinect Body Tracking SDK
- OpenCV
See the cwipc readme file for installation instructions.
The k4a capturer needs a cameraconfig.xml file that specifies the serial numbers of the cameras and their position and orientation.
If you have a single camera in landscape mode looking forward, approximately 1m from the subject at 1m height you can automatically create a config file with
cwipc_calibrate --kinect --auto
If you have more cameras, or a single camera that is oriented differently (for example in portrait mode, so it is easier to capture a full human body) you need to print a target from ../cwipc_util/target-a4.pdf. Easiest is to put the target on the floor, at the position that will become (0, 0, 0) in your captured pointclouds. Then run
cwipc_calibrate --kinect --target a4floor --nofine
It is possible to hold the target in a different position, or use different targets, depending on the physical layout of your space. Use
cwipc_calibrate --listto see the options.
You now inspect the frame captured to check that the target is fully visible in all cameras. If it is not you reorient and restart the calibration.
Next you select the 4 colored points on the virtual target, remembering the colors.
Then, for each of the cameras, you select the corners of the A4 paper with the same colors in the same order, and inspect the single-camera result.
Finally you inspect the merged pointcloud.
This gives you a course calibration in cameraconfig.xml.
Inspect the resulting pointcloud view with
cwipc_view
If you want to fine-tune calibration you can grab a pointcloud with the w key. You can then use
cwipc_view --nograb file.ply --reuse --nocoarse
to attempt fine calibration but this is a bit of a black art.
You can edit cameraconfig.xml and modify various settings, such as camera parameters like white balance, various processing options that govern how RGB and Depth images are converted to pointclouds, and bounding box parameters for the pointcloud.
After editing parameters you re-run cwipc_view to see the effect of your changes.