研究目的
To present a method for modeling the intrinsic and extrinsic parameters of the infrared and colour cameras, and more importantly the distortions in the depth image of the Kinect system, through an integrated marker- and feature-based self-calibration.
研究成果
A self-calibration method suitable for the Microsoft Kinect was presented and tested. The method solves for the relative translations and rotations between the IR camera, projector, and RGB camera. At the same time, it solves for the intrinsic parameters of both cameras, extrinisic parameters of the IR camera, object space target coordinates, and plane parameters. Geometric constraints have been included in the bundle adjustment to ensure points lie on the best fit plane and the optical sensors are all mounted rigidly on the same platform. The depth calibration is expressed as a function of three rotations, three translations, interior orientation parameters and the lens distortion of the IR camera.
研究不足
The calibration model does not characterize the depth measurements of the Kinect as a function of the IOPs and APs of both sensors and the six ROPs defining the stereo pair. The effect of quantization of the disparity measurements is included in the stochastic model despite its small effect on reconstruction accuracy.
1：Experimental Design and Method Selection
The methodology involves a bundle adjustment with self-calibration model to solve for the intrinsic and extrinsic parameters of the infrared and colour cameras, and the distortions in the depth image. The model includes geometric constraints to ensure points lie on the best fit plane and the optical sensors are all mounted rigidly on the same platform.
2：Sample Selection and Data Sources
Two Kinect for Xbox sensors were used. Images were captured using the standard VGA resolution to ensure that the IR images are calibrated at the same image resolution as the depth images.
3：List of Experimental Equipment and Materials
Microsoft Kinect for Xbox sensors, planar checkerboard target, FARO Focus3D terrestrial laser scanner.
4：Experimental Procedures and Operational Workflow
Following a two hour warm-up period, depth images, IR images, and RGB images of a checkerboard target were acquired from various positions and orientations. At every exposure, 20 consecutive depth images were captured and averaged to reduce the random noise of the depth measurements and to fill in holes in the depth map.
5：Data Analysis Methods
The collinearity and coplanarity equations are highly non-linear so the Gauss-Helmert least-square model has been chosen for minimizing the summation of the weighted residuals. Baarda’s data snooping with a 5% level of significance was used to minimize the possibility of outliers in the adjustment.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
