Abstract: An algorithm of weld seam feature detection was presented, which was applied to stereo matching of stereo vision system for remote welding.Firstly, the bottom-hat transformation in grey-scale morphology was used to preprocess the image, making the weld seam section more distinctive, and Canny edge detection and close operation were employed to obtain a pixel level closed edge of weld seam;secondly, the edge data was filtered according to continuity and limited width of the weld seam edge;finally, the sub-pixel weld seam edge data was gained by fitting a cubic smoothing spline to the filtered edge data.With the detected sub-pixel weld seam, stereo matching and 3D reconstruction experiments were performed. The results show that the algorithm is rapid, exact and robust to noises.