Stereo


The stereo system used by Ratler takes its input from two black-and-white CCD cameras with auto-iris, 8 mm lenses, mounted on a motion-averaging mast. The field of view is about 50 degrees horizontal and 42 degrees vertical. The output is sets of (x,y,z) triples, given in the camera coordinate frame, along with the pose of the robot at the time the images were acquired. Using the pose, the (x,y,z) values can be transformed into world coordinates to form a (non-uniformly distributed) terrain elevation map which is used by the planning software to negotate the terrain. The stereo system can be requested to process only part of the image, and that at reduced resolution (skipping rows and columns in the image), to speed up the overall system cycle time.

The stereo images are first rectified to ensure that the scan lines of the image are the epipolar lines [Robert et al. 1994a]. The best disparity match, within a given window, is then computed using a normalized correlation. Disparity resolution is increased by interpolating the correlation values of the two closest disparities. The normalized correlation method is relatively robust with respect to differences in exposure between the two images, and can be used to produce confidence measures in the disparity values.

Care must be taken to ensure that outlier values (caused by false stereo matches) are minimized. Several methods are used to achieve the level of reliability required for navigation. One method is to have lower bounds on the acceptable correlation values and variance in pixel intensity, to eliminate low-textured areas. Another method eliminates ambiguous matches (caused by occlusion boundaries or repetitive patterns) by rejecting matches that are not significantly better than other potential matches. Finally, the values can be smoothed to reduce the effect of noise. All these methods help to produce elevation maps that accurately reflect the actual surrounding terrain, with only a few centimeters of error.


LRD Navigation Group - skoenig@cs.cmu.edu (last updated in March 1995)
Comments? Suggestions? Requests? Please send e-mail to lri-feedback+@cs.cmu.edu!