next up previous
Next: Stereo computation Up: Stereo Previous: Stereo

Camera Calibration

 

The internal parameters of each camera are computed using an extension to the Tsai method[8] that employs a nonlinear gradient descent technique to compute the centre of radial distortion. This provides a model of the radial distortion and focal length of the lenses. The radial distortion is quite significant because the lenses' 80 degree field of view is quite wide. A warp table is generated that maps the original images into subsampled versions that have been corrected for lens distortion and have epipolar lines that are aligned with the x and y axis of the image. An example input image is shown in Figure 3, and the undistorted image is shown in Figure 4.

The position of each camera is computed relative to a calibration target. This information is used to compute the transformations (rotation and shear) needed to align the upper and left cameras with the right camera. An accurate calibration is important in obtaining stereo matches with little noise and in projecting the stereo data into 3--D coordinates.

  
Figure 3: Right Image

  
Figure 4: Corrected Image

The robot is calibrated several times at different angles of rotation, to determine the centre of rotation for the robot. This allows the stereo depth maps to be projected into volumes in 3--D space in a coordinate system relative to the center of the robot.



Vladimir Tucakov
Wed Dec 4 11:45:59 PST 1996