Aiming at the problem of how to quickly and accurately obtain complete 3D point cloud data within a mold, a 3D point cloud stitching algorithm combining a robotic arm and a 3D visual device is proposed. In the initial stitching process, the hand-eye calibration method is used to obtain the hand-eye matrix, and the point cloud is converted to the robot base coordinate system to complete preliminary stitching and obtain a good initial registration position. On this basis, an improved iterative closest point (ICP) algorithm is proposed, which combines intrinsic shape signature (ISS) features to obtain key points, and uses random consistency algorithm to eliminate erroneous matching points. During the point matching process, point to face matching is used to obtain a complete spliced point cloud. The experimental results show that the proposed algorithm has good robustness in comparison with the ICP algorithm and other improved ICP algorithms in terms of registration performance, and the algorithm time consumption and registration error are significantly reduced. The error of a complete mosaic point cloud is 0.12 mm, which has high engineering practical value.