基於幾何特徵保留的自適應光學點雲簡化

Abstract

隨著攝影測量技術的普及與應用領域擴展，三維點雲資料量快速增加，如何在有效簡化點數的同時保留關鍵幾何特徵，成為點雲處理中的重要課題。現有多數簡化方法倚賴固定參數或經驗法則，對不同區域特性之適應性不足，容易導致幾何細節流失或過度簡化。 本研究提出一種改良式點雲簡化策略，依據邊緣點、特徵點與非特徵點的分類進行分層處理，並透過自適應鄰域進行特徵保留。首先，對原始點雲進行雜訊排除，並結合主成分分析計算曲率與熵，以建構點雲拓樸特徵並劃分為散亂與規則區域。接續進行邊緣點萃取與特徵點重要性分析，再依據該資訊於非特徵區域中進行簡化，最終達成兼顧特徵保留與點雲減點的目的。 本研究於五組模擬資料集（Bunny、Dragon、Armadillo、Happy Buddha、Asian Dragon）並與三種既有方法（On Graph、Feature Fusion、MIWSA）比較。結果顯示，在不同簡化率下，本方法於特徵保留能力、幾何誤差控制與體積穩定性方面皆具優勢。另外也對四組光學點雲（兩組建築、雕像、浮雕等）進行簡化測試，結果顯示在細節豐富或結構複雜之點雲中，仍能保留清晰輪廓與幾何完整性。 實驗結果顯示，本方法在多樣化場景下皆具備良好的幾何特徵保留能力，即便在高簡化率條件下，仍能維持整體形體完整性。然而，由於非特徵區域在簡化過程中點數相對較少，可能導致局部誤差略高。除此之外，為進一步提升方法效能與穩定性，未來可考慮納入更多參數自動化調整機制，以增強不同場景下的適應性。

With the increasing popularity of photogrammetry and the expanding scope of its applications, the volume of 3D point cloud data has grown rapidly. Effectively reducing point count while preserving key geometric features has become a critical challenge in point cloud processing. Most existing simplification methods rely on fixed parameters or empirical rules, which often lack adaptability to varying regional characteristics, leading to the loss of geometric details or excessive simplification. This study proposes an improved point cloud simplification strategy based on the classification of edge points, feature points, and non-feature points, with a hierarchical processing approach and adaptive neighborhood for feature preservation. The process begins with noise removal from the original point cloud, followed by curvature and entropy estimation via principal component analysis to construct topological features and segment the data into scatter and regular regions. Subsequently, edge points are extracted, and feature point importance is analyzed. Simplification is then performed in non-feature regions based on this information, ultimately achieving a balance between feature preservation and point cloud simplification. The proposed method is evaluated on five simulated datasets（Bunny, Dragon, Armadillo, Happy Buddha, and Asian Dragon）and compared with three existing methods （On Graph, Feature Fusion, and MIWSA）. Results demonstrate that across different simplification ratios, the proposed method exhibits superior performance in preserving features, controlling geometric error, and maintaining volume stability. Additionally, tests conducted on four photogrammetric point clouds （including two sets of buildings, statue, and bas-relief）show that even for detail-rich or structurally complex data, the method effectively retains clear contours and geometric integrity. Experimental results confirm that the proposed approach consistently maintains geometric features across diverse scenarios, even under high simplification ratios. However, due to the reduced point count in non-feature regions during simplification, some localized errors may occur. In addition, to further enhance the method’s efficiency and robustness, incorporating automated parameter tuning mechanisms is suggested to improve adaptability across various applications.