Edge chaining using smoothly-varying stereo disparity

Claim 1

A method for linking edges in stereo images into chains which correspond to physical contours in a scene, comprising: detecting edges in the stereo images; matching edges between the stereo images, such that each edge has associated stereo matches; linking the edges into chains, such that each chained edge has a single associated stereo match, and neighboring edges in a chain have matches with a substantial similar stereo disparity; and in which the stereo images are acquired with more than two cameras, and further comprising: identifying one stereo image as a primary image and the remaining stereo images as secondary images; determining, for each edge in the primary image, an epipolar line relative to each of the secondary images; determining, for each such epipolar line, a difference in angle between an edge alignment and the epipolar line, and finding a maximum difference in angle over all the secondary images, and identifying the secondary image which is associated with a maximum difference as a target image; identifying the remaining secondary images as support images; measuring, for each edge in the primary image and the edges along the epipolar line in the target image, a difference in gradient direction, and rejecting any edges as matching if the difference is greater than a first threshold; otherwise measuring, for each edge in the primary image and the edges along the epipolar line in the target image, a difference between intensities of pixels neighboring the edges, and rejecting the edges as matching if the difference is greater than a second threshold, otherwise rejecting any edges as matching if there is no corresponding edge in the support images; and otherwise accepting the edges as matching.

Claim 2

A method for linking edges in stereo images into chains which correspond to physical contours in a scene, comprising: detecting edges in the stereo images; matching edges between the stereo images, such that each edge has associated stereo matches; linking the edges into chains, such that each chained edge has a single associated stereo match, and neighboring edges in a chain have matches with a substantial similar stereo disparity; and in which the stereo images are acquired with more than two cameras, and further comprising: identifying one stereo image as a primary image and the remaining stereo images as secondary images; determining, for each edge in the primary image, an epipolar line relative to each of the secondary images; determining, for each such epipolar line, a difference in angle between an edge alignment and the epipolar line, and finding a maximum difference in angle over all the secondary images, and identifying the secondary image which is associated with a maximum difference as a target image; identifying the remaining secondary images as support images; measuring, for each edge in the primary image and the edges along the epipolar line in the target image, a difference in gradient direction, and rejecting any edges as matching if the difference is greater than a first threshold; otherwise measuring, for each edge in the primary image and the edges along the epipolar line in the target image, a difference between intensities of pixels neighboring the edges, and rejecting the edges as matching if the difference is greater than a second threshold, otherwise rejecting any edges as matching if there is no corresponding edge in the support images; and otherwise accepting the edges as matching.