The robot teaching data can be classified by the shape of the weld seam. From the 3-D coordinates of the weld seam, the shape of the weld seam and the torch orientation can be computed by using the stereo vision algorithm. In these processes, the correspondency is achieved by the correlation technique or using the intersecting points of epipolar lines and edges. The normal vector of the plane is computed by crossproducting the direction vectors of each straight lines.
The experimental result shows that the arcs with various radius converge and the accuracy of the 3-D coordinates are within the 3 mm, and the error on the normal vector of the plane is within 6.