Game Development Reference
In-Depth Information
image in real-time by using area based stereo matching (Konolige, 1997). Their
algorithm has four major blocks, namely LOG transform, variable disparity
search, post-filtering, and interpolation. The special purpose hardware consists
of two CMOS 320x240 grayscale imagers and lenses, low power A/D convert-
ers, a digital signal processor, and a small flash memory for program storage. The
board communicates with the host PC via the parallel port. Second generation
hardware uses a DSP from Texas Instruments (TMS320C60x).
Reconstruction and calibration
Reconstruction involves computing a point in space for each corresponding point
pair in the images. This requires calibration of the cameras. There are two major
parameter sets for cameras, namely intrinsic and extrinsic parameters. If both
of the parameter sets are known, then the cameras are fully calibrated. By using
the intrinsic parameters, the 3D depth map can be converted into (x,y,z)
coordinates. The depth values give the z-coordinates and (x,y) coordinates are
calculated from camera's intrinsic parameters. The extrinsic parameters are
used to convert the camera centered (x,y,z) position into a world coordinates
position (Narayanan et al., 1998; Kanade et al., 1997). These 3D points are
converted into a surface representation via a triangular mesh. Since there is no
exact solution, the algorithm calculates the correspondence that minimizes the
geometric error subject to the epipolar constraint. In this chapter, for our
experiments we assume that the cameras are fully calibrated. Detailed informa-
tion about cameras and camera calibration can be found in Hartley & Zisserman's
work (Hartley, 2000).
An exemplar application for scene reconstruction is Narayanan et al.'s (1998)
work. The authors use 51 synchronized and calibrated video cameras to extract
the depth map, polygonize it into triangles in 3D space, and apply texture maps
onto the mesh. Another 3D scene reconstruction method is volumetric recon-
struction. In this method, the reconstruction volume is divided into voxels where
volumetric intersection algorithms reconstruct surface and voxels from the
silhouette of an object (Cheung et al., 2000).
Pollefeys et al. (1999) developed a structure from the motion method to
reconstruct a scene from uncalibrated cameras. Structure from motion was also
used by Zisserman et al. (1999) for scene reconstruction. In their method, the
authors locate corners in the images and estimate the fundamental matrix.
Although many algorithms are proposed for more accurate and reliable 3D object
reconstruction, they are not suitable for practical applications due to their
computational complexity. Depending on the application type, algorithm and
hardware-related solutions are proposed. In Li et al. (2001), the authors reduce
the complexity of finding spatio-temporal correspondence by using constraints
Search WWH ::
Custom Search