To enable intelligent robotic welding, this paper proposes a method for attitude recognition and weld tracking tailored for spatial right-angle welds. A combination of second-order difference method based on point cloud contours and least-squares linear fitting is used to extract welding coordinates from a single-frame laser point cloud. During the welding process, these identified coordinates are sequentially stored in a queue to form a continuous weld trajectory. To refine the trajectory, a smoothing method comprising limiting processing and cubic B-spline curve fitting algorithm is applied. Tangent vectors along the trajectory are then used to construct a computational model for weld positioning. By leveraging the weld points and attitude vectors, the robot dynamically adjusts its position in real time, ensuring accurate alignment and execution of the welding process. Experimental validation con-firms that the root mean square error between measured and actual points remains within 0.26 mm, meeting the precision requirements for high-quality welding. This approach demonstrates a robust solution for automated welding in complex scenarios.