Multi-Kinect camera calibration

Pairwise and octagonal calibration methods for eight Kinect v2 sensors. CVIP 2021.

NIT Rourkela, May 2018 to Jan 2020

Developed two methods for calibrating eight Microsoft Kinect v2 sensors, as the 3D capture setup for my undergraduate thesis. The first is pairwise calibration: consecutive pairs of Kinects view a 3D object on a treadmill, and calibration runs in six steps, from pre-calibration and image capture through exporting camera parameters, generating live point clouds and matching them, to checking the merged point cloud. The second is simultaneous single-camera calibration of all eight cameras on the same setup with a novel octagonal model. The two methods reached overall mean reprojection errors of 1.37 and 0.42 pixels, so the octagonal method cut the error by 69%, with best cases of 0.63 and 0.27 pixels, and outperformed existing techniques on both root mean square and reprojection error (Hazra et al., 2021).

This built on earlier work in the summer of 2018 in the Machine Intelligence and Bio-Motion Laboratory, extracting people from Kinect v2 video with adaptive background subtraction.

References

2021

  1. CVIP 2021
    Two Novel Methods for Multiple Kinect v2 Sensor Calibration
    Sumit Hazra, Manasa Pisipati†, Amrit Puhan†, and 2 more authors
    In International Conference on Computer Vision and Image Processing (CVIP), 2021