Robust 3D Point Cloud Registration Based on Deep Learning and Optimization

Publication Type:
Issue Date:
Full metadata record
Most existing learning-based point cloud matching methods suffer one or more of the following limitations: (1) depending on supervised information from manually labeled data, which is tedious and labor-intensive, (2) suffering from performance degradation to handle point-cloud pairs with large rotation, partial overlaps, and density variations, (3) without integrating point and structure matchings into one stage for searching correspondences, and correspondences are obtained by nearest neighbor search (NN) of local feature descriptors, resulting in high outlier rates, (4) relying on attention mechanisms to simulate soft matching with high computation and memory cost, mainly when being applied to points of a larger number. This thesis is conducted using optimization theory and deep learning techniques to alleviate the limitations mentioned above. For one thing, it aims to develop unsupervised methods for learning feature representations of point clouds to reduce reliance on human annotations. Another focus of this thesis is on formulating optimization techniques to address registration tasks involving large rotations effectively. Finally, it is also dedicated to designing algorithms to establish more accurate correspondences for point cloud registration with low partial overlaps and density variations. To fulfill these goals, Chapter 3 proposes a soft clustering-based unsupervised algorithm to learn distinctive point cloud representations. The proposed method does not depend on data augmentation, which differs from previous unsupervised works. Chapter 4 extends a correspondence-free method to solve point cloud matching with large rotations and partial overlaps. Expressly, a rotation-based unsupervised method is first provided to learn rotation-sensitive features. Then, a beam search-based scheme incorporates networks to initialize correspondence-free methods to solve large rotation registration. Finally, a clustering-based soft-segmentation approach is employed to solve point cloud alignment with partial overlaps. Chapter 5 develops a fused optimal transport-based algorithm to establish more correct correspondences in the detected overlapping regions for solving partial overlap registration by integrating point and structure matching. Chapter 6 integrates the overlap scores into a probabilistic registration method to cope with point cloud registration with partial overlaps and density variations. Furthermore, it also provides a clustering-based attention model that simulates matching with low computation and memory cost. All the proposed point cloud matching methods are evaluated on many registration benchmarks, showing their potential to contribute to registration development.
Please use this identifier to cite or link to this item: