Abstract: In view of the low accuracy, efficiency and robustness of traditional point cloud registration in complex scenes, which are difficult to cooperate with robots in industrial operations, a point cloud registration method based on signature of histograms of orientations (SHOT) descriptor was proposed. The model point cloud and field point cloud were pre-processed by downsampling the center of gravity of voxels, and the internal shape signature (ISS) was used to extract feature points from the downsampled point cloud. The SHOT descriptors of feature points were calculated, and the KD tree was constructed to quickly retrieve the pairs of points with similar features. The random sampling consistency (RANSAC) algorithm was used to remove the mis-matched point pairs and complete the coarse alignment, obtain the initial position of the coarse alignment of the point cloud, and complete the fine alignment by combining with iterative closest point (ICP). The experimental results show that this method can quickly identify and locate H-beams and thermocouples in complex scenes, with an alignment time of 2 s and an alignment accuracy within 5 mm. Compared with the traditional ICP algorithm, this method has higher alignment accuracy and better robustness, this method can identify point clouds of objects with a certain degree of occlusion and mutilation, and has better robustness, which can meet industrial needs.