Hierarchical Cross-Source Point Cloud Registration Method Based on Adaptive Instance Segmentation

During the movement of mobile agents, high-precision pose information is obtained by fusing data from different sensors through cross-source point cloud registration. However, challenges such as density differences between modalities and low field-of-view overlap are encountered in this process. To address the issue that traditional optimization or deep learning methods struggle to balance global consistency and local accuracy in complex multi-object environments, a hierarchical method based on adaptive instance segmentation (AIS-HCSR) is proposed. This method constructs a scene-object-point cloud hierarchical progressive registration framework: Firstly, at the scene level, it fuses distance and angle features through an adaptive geometric feature encoding mechanism and dynamically adjusts feature weights based on local geometric complexity to achieve initial matching across source scenes. Then, at the object level, an adaptive Euclidean clustering algorithm is introduced for point cloud instance segmentation, and an instance correspondence mechanism based on superpoint matching propagation and spatial layout consistency verification is designed to eliminate matching ambiguities in the relative positions of objects. Finally, at the point cloud level, object-level point-plane residual optimization and global modality consistency fusion are combined to achieve fine registration that retains local accuracy while ensuring global consistency.This hierarchical strategy effectively addresses the challenges of complex scenes with large density differences, partial overlaps, and similar multiple objects by integrating macroscopic structural information of the scene with microscopic geometric features of the objects, thereby enhancing the accuracy and robustness of cross-source point cloud registration.