一種快速且穩健的口內掃描點雲配準方法

Abstract

本論文提出一種快速且具迭代性的深度學習式點雲配準流程，應用於手持式高速結構光口內掃描儀之全口重建。傳統之迭代最近點（Iterative Closest Point, ICP）配準方法對初始位姿高度敏感，於部分重疊或含雜訊之掃描資料中容易陷入區域最小值；而近年之深度學習特徵式配準方法多仰賴高計算成本之密集對應搜尋，限制其於臨床即時應用的可行性。此外，口內掃描資料常包含大面積平坦之軟組織區域（如牙齦或齒槽嵴），其局部幾何特徵不明顯，使得基於局部特徵之配準方法可靠度降低。 為克服上述問題，本文方法不進行任何顯式特徵描述子比對，而是整合三個階段之由粗至細配準架構：（一）以體素化為基礎之重疊區域萃取，以降低資料規模；（二）利用 Lucas–Kanade（LK）最佳化架構進行全域粗配準，以提升初始對齊之穩健性；以及（三）透過迭代最近點法進行細部幾何之精細配準。藉由先行體素化並萃取潛在重疊區域，本方法在執行 LK 配準前有效減少點雲數量，避免大規模 RANSAC 或特徵比對所帶來之高計算負擔，並實現快速且穩定之收斂。 實驗結果顯示，在高重疊條件下，所提出的方法於真實口內掃描資料及合成之 ModelNet40 資料集上，於配準精度與計算效率方面皆優於多種現有先進方法，顯示其具備應用於臨床即時全口重建之潛力。

We present a fast, iterative deep learning (DL) based registration pipeline for full-mouth reconstruction using handheld high-speed structured-light intraoral scanners. Traditional Iterative Closest Point (ICP) registration methods require good initialization and can easily become trapped in local minima under partial overlap or noisy scans. Recent DL feature-based registration frameworks rely on a computationally expensive dense correspondence search, which limits their utility in clinical settings. Additionally, intraoral scans often contain large planar soft-tissue regions such as the gingiva or alveolar ridge, which lack distinctive local geometric features, making local feature-based registration less reliable. To address this, our approach omits any explicit descriptor matching by combining (1) voxel-based extraction to isolate potentially overlapping regions, (2) a Lucas–Kanade (LK) based global registration for robust coarse alignment, and (3) a final Iterative Closest Point (ICP) refinement to recover fine dental geometry. By first voxelizing and extracting, we reduce the data size before invoking LK, resulting in rapid convergence without large-scale RANSAC or feature matching. Under high overlap conditions, our method outperforms state-of-the-art approaches on both real intraoral scans and the synthetic ModelNet40 dataset, demonstrating both speed and accuracy for chairside use.