Automatic welding of saddle-shaped welds has the problems of difficult identification and localization of the weld seam, and difficult planning of welding torch posture. In order to realize the automatic welding of saddle-shaped welds, a saddle weld identification and planning method based on point cloudis proposed using the actual assembly of pipe cross structures as the research object. This method utilizes tight coupling constraints between the normal vector and spatial distance to extract the key pointsaccording to the spatial semantic features. And the key points are approximated by three-times cubic B-spline curvesto obtain the final spatial welding curve. Experiments show that the maximum error of the extracted weld trajectory compared with the real trajectory is 0.68 mm, and the maximum root mean square error is 0.29. The method has high accuracy and robustness, which can meet the needs of practical welding. The practicality of this method is not only limited to saddle-shaped weld seams, but also can deal with other weld seams with curvilinear variations.