Per-Pixel Calibration for RGB-Depth Natural 3D Reconstruction on GPU

OF THESIS Per-Pixel Calibration for RGB-Depth Natural 3D Reconstruction on GPU Ever since the Kinect brought low-cost depth cameras into consumer market, great interest has been invigorated into Red-Green-Blue-Depth (RGBD) sensors. Without calibration, a RGBD camera’s horizontal and vertical field of view (FoV) could help generate 3D reconstruction in camera space naturally on graphics processing unit (GPU), which however is badly deformed by the lens distortions and imperfect depth resolution (depth distortion). The camera’s calibration based on a pinhole-camera model and a high-order distortion removal model requires a lot of calculations in the fragment shader. In order to get rid of both the lens distortion and the depth distortion while still be able to do simple calculations in the GPU fragment shader, a novel per-pixel calibration method with look-up table based 3D reconstruction in real-time is proposed, using a rail calibration system. This rail calibration system offers possibilities of collecting infinite calibrating points of dense distributions that can cover all pixels in a sensor, such that not only lens distortions, but depth distortion can also be handled by a per-pixel D to ZW mapping. Instead of utilizing the traditional pinhole camera model, two polynomial mapping models are employed. One is a two-dimensional high-order polynomial mapping from R/C to XW/YW respectively, which handles lens distortions; and the other one is a per-pixel linear mapping from D to ZW , which can handle depth distortion. With only six parameters and three linear equations in the fragment shader, the undistorted 3D world coordinates (XW , YW , ZW ) for every single pixel could be generated in real-time. The per-pixel calibration method could be applied universally on any RGBD cameras. With the alignment of RGB values using a pinhole camera matrix, it could even work on a combination of a random Depth sensor and a random RGB sensor.