Abstract: Objective Aiming at technical bottleneck of difficult identification of saddle-shaped weld groove features of TKY tubular joints in offshore oil and gas platforms, an intelligent detection algorithm for dynamic segmentation of three-dimensional point cloud based on laser vision is proposed to realize high-precision identification of geometric features of weld grooves. Methods The high-precision point cloud data of saddle-shaped weld groove was obtained by line laser vision sensing system. Based on differential geometric statistical characteristics, combined with the improved three-dimensional point cloud change point detection algorithm, dynamic segmentation of multi-geometric feature regions such as outer wall of branch pipe, side wall of groove, backing welding and outer wall of the main pipe was realized. Combined with the first-order differential continuity criterion and the second-order differential characteristic analysis, weld feature points were accurately positioned, and quantitative models of key parameters such as weld bead backing width, penetration and groove opening were established. Results Experiments show that this method can accurately identify feature points of saddle-shaped weld groove, accurately measure geometric size of weld, and realize high-precision feature extraction of complex geometry of branch pipe and main pipe, and provide data support for autonomous planning of multi-layer and multi-pass welding of the robot. Conclusion The laser vision three-dimensional point cloud dynamic segmentation detection algorithm can significantly improve recognition accuracy of weld features of thick-walled complex structures, provide a technical basis for the development of weld automatic welding technology, and has good engineering application prospects.