A modified feature based stereo vision [sic] technique is described in this thesis. The technique uses the curve-segments as the feature primitives in the matching process. The local characteristics of the curve-segments are extracted by the, Generalized Hough Transform.

A set of images of a scene, which are taken by a multi-positioned camera satisfying the parallelism criterion, are first filtered by the Laplacian of a Gaussian operator in different widths, i.e. coarse to fine channels. At each channel, the Generalized Hough Transform is applied to the curve-segments in each image. The curve position, the curve-length, the curve centroid, the average gradient of the curve-segment and the R-table are, used as the local features in representing the distinctive characteristics of the curve-segment. These features of all the curve-segments, in an image are used as the constraints to find the corresponding curve-segments in the different images. The epipolar constraint oil the centroid of the curve-segment is used to limit the search window ill the images. Since the multi-images of one view are used, there exist more information about the scene than the two relational images criterion. Its performance compares to the other matching techniques, for example, the point, matching or twin image matching that the mismatching and the calculation are greatly reduced. Although the algorithm is not feasible for the realization of the real-time implementation of stereo vision [sic], it is a more economic way of finding the depth of all object or a view.