張鈞法Chang, Chun-Fa陳建壹Chen, Chien-Yi2025-12-092025-07-312025https://etds.lib.ntnu.edu.tw/thesis/detail/4563917caae413ce01ecf91eec1b7abb/http://rportal.lib.ntnu.edu.tw/handle/20.500.12235/125833樹木是自然環境中不可或缺的元素，不僅是環境重建與山林模擬的重要目標，也廣泛應用於遊戲、動畫與電影場景中，能有效營造氛圍並提升沉浸感。然而樹木具有複雜的分枝結構與茂密的樹冠，傳統人工建模方式不僅耗時費力，也難以應對大量模型的產出需求。因此發展一種兼具效率與準確性的建模方法，有助於建立具備實用性與可重複利用價值的樹木模型資產，滿足大規模生成與多元應用的需求。本研究提出一套結合運動回復結構技術（Structure-from-Motion, SFM）與 L-system 的程序化建模方法。首先透過多視角影像重建點雲，擷取樹木區域並濾除無效點，接著應用 AdTree 演算法提取骨架結構，建立初步樹狀模型。考量 AdTree 對雜訊與葉片分布的敏感性，可能造成末端分枝過密與不自然，本研究進一步透過規則修剪異常結構，並引入 L-system 根據樹種特性生成自然的細枝與葉片分布，以提升模型的結構合理性與視覺自然感。此外針對因視角不足造成的點雲缺失，亦可將現有分支轉換為可重用的 L-system 子樹，由使用者選擇節點進行接合補齊。最終將完整的樹木結構輸出為可編輯的三維網格模型（如 .obj），可以應用於各類虛擬場景，展現本方法於樹木建模上的靈活性與實用性。Trees are essential elements in natural environments, playing crucial roles not only in environmental reconstruction and forest simulation but also widely in games, animations, and film scenes, effectively creating atmosphere and enhancing immersion. However, trees have intricate branching structures and dense canopies, making traditional manual modeling methods both labor-intensive and inefficient for producing a large volume of models. Consequently, developing a modeling approach that balances efficiency with accuracy would significantly benefit the creation of practical and reusable tree model assets, fulfilling the requirements of large-scale generation and diverse applications. This study proposes a procedural tree modeling method combining Structure-from-Motion (SFM) technology and L-systems. First, point clouds are reconstructed from multi-view images, and irrelevant points are filtered out to isolate tree regions. Then, the AdTree algorithm is applied to extract initial skeletal structures and construct preliminary tree models. Considering the sensitivity of AdTree to noise and uneven leaf distribution—which can result in excessively dense or unnatural end branches—this research further employs rule-based pruning to eliminate anomalous structures. Additionally, the L-system is introduced to generate natural fine branches and leaf distributions based on specific tree characteristics, improving structural rationality and visual realism. Moreover, to address missing point cloud data due to insufficient viewpoints, existing branches can be converted into reusable L-system subtrees, allowing users to select nodes for seamless integration and completion. Finally, the complete tree structure is exported as editable three-dimensional mesh models (e.g., .obj), suitable for various virtual scenarios, demonstrating the flexibility and practicality of this modeling method.點雲L-system三維重建樹木重建運動回復結構電腦圖學Point CloudL-system3D ReconstructionTree- ModelingStructure-from MotionComputer Graphics學術論文