Robust multi-view registration method for narrow scenes

Multi-view point cloud registration is one of the key steps in reverse engineering, which has important research significance and engineering application value. As for point cloud data obtained from narrow scenes (such as oral cavity or mechanical structure), the accuracy of the multi-view registration algorithm directly affects the accuracy of the reconstructed results. In order to improve the speed and robustness of multi-view registration for narrow scenes, an incremental multi-view point cloud registration method based on pose optimization is proposed. Firstly, a multi-strategy registration algorithm is proposed based on iterative closest point method (ICP) and feature-based registration method to solve the registration of adjacent point clouds. Then, based on the incremental registration of adjacent point clouds, a loop closure detection method based on distance constraints is proposed, and the pose graph is constructed according to the registration results of adjacent point clouds and loop closure detection results. Finally, the real-time optimization strategy is used to optimize the pose graph to alleviate drift errors and achieve robust multi-view registration. Experimental results show that the proposed multi-strategy registration algorithm and the loop closure detection method with distance constraints are effective. The classical ICP algorithm and the FPFH-based method are invalid in the experiment, but the proposed multi-strategy registration algorithm is valid. The loop closure detection method with distance constraints is more efficient than the conventional loop closure detection method. The multi-view registration algorithm proposed in this paper can achieve accuracy of 0.0357 mm in tooth model data registration. In order to verify the universality of the algorithm, the model point clouds collected continuously in multiple narrow scenes are used for verification. The results show that the proposed algorithm achieves good results, which indicates that the proposed method is an effective multi-view registration method for narrow scenes.