This dissertation addresses the challenges of robust shape modeling and multi-object segmentation for non-cooperative proximity operations using 3D point clouds. The method targets scenarios in which only noisy, partially observed point measurements a...
This dissertation addresses the challenges of robust shape modeling and multi-object segmentation for non-cooperative proximity operations using 3D point clouds. The method targets scenarios in which only noisy, partially observed point measurements are available, and the resulting geometric representation must be both expressive for complex shapes and compact enough to serve as an analytic constraint for downstream applications. To this end, a closed-form implicit surface modeling framework based on even-degree homogeneous polynomials is developed. The target boundary is represented as a polynomial level set, enabling consistent inside/outside evaluation with a single constraint-form geometry.
To improve numerical stability and mitigate overfitting associated with high-order polynomial fitting, a Fischer-decomposition-based orthogonal basis is introduced. This representation provides rotation-compatible, mutually orthogonal components and supports controlled model complexity through selective use of dominant harmonic terms. For robustness against outliers and sensor noise, a RANSAC-driven estimation pipeline is integrated with adaptive reweighting to refine the final model based on residual-consistent inliers.
For multi-object scenes, a graph-based segmentation strategy is incorporated. A proximity graph constructed from the point cloud is partitioned via spectral clustering, and the proposed polynomial fitting procedure is applied to each cluster independently.
Experimental validation demonstrated that the 4th order polynomial achieves optimal performance with SDF errors of approximately 5% of the obstacle size.
Overall, the proposed approach provides an analytic, high-fidelity implicit representation directly from raw 3D point clouds and establishes a practical pathway for integrating point-cloud-derived obstacle models into constraint-based proximity-operation algorithms.