An overall visual algorithm system for sub-assembly welding robots is established based on the point cloud data of the sub-assembly complex structural components captured by 3D cameras. When the point cloud images are processed, it is found that the shadows and conventional point cloud filtering methods may damage the model, so the outlier filtering algorithm is employed to clean the point cloud data, followed by the estimation of normal vectors based on extracted features, and the small-scale shadow noise planes are removed combined with the curvature filtering algorithm. Through equipment testing and production validation, the noise points are successfully eliminated and the sharp features of the 3D model are maintained, setting the stage for subsequent point cloud registration and 3D reconstruction. The entire sub-assembly 3D camera processing system can cover most complex structural components of the sub-assembly. The findings of this study offer valuable insights for the design of a visual system for vertical weld identification in shipyard sub-assembly teams.