Compact 3D building models from mobile laser scanning
本文以移動機器人上的雷射測距儀建立室內外建物的精簡 3D 模型。室內機器人以水平雷射做 2D 掃描匹配求位姿，同時以朝上的雷射掃出 3D 結構；室外機器人則以裝在雲台上的單一雷射取得 3D 掃描，並以射線式機率模型對三角網格做 3D 配準。取得的網格全域一致但局部雜訊大，作者以隨機起點的區域成長找出大型平面，把點投影到平面後合併共面多邊形，同時保留門窗等非平面細節；為加速，先由各掃描抽出的線段角度直方圖產生平面候選，再以粗細兩階段的平面掃掠篩選。
本頁內容
Mobile laser 3D modelling that estimates poses by 2D and probabilistic 3D scan matching and then replaces noisy meshes with large planar polygons found by randomized region growing and histogram-guided plane sweeps.
技術屬性
欄位內容為文獻擷取紀錄的原文用語（英文），以原文為據；「未查證」表示本研究尚未讀到該資訊，不代表該方法不具備此能力。
| 感測輸入 | 2D laser range finders: indoors a horizontal laser for 2D mapping plus an upward-pointing laser for 3D (SICK PLS used in the Wean Hall run; SICK LMS also named); outdoors one laser on a pan/tilt unit、odometry |
|---|---|
| 原文測試平台 | wheeled UGV (indoor robot with two lasers)、wheeled UGV (outdoor robot Herbert with a pan/tilt laser) |
| 狀態估計 | incremental maximum-likelihood pose estimation by hill climbing: 2D scan-to-grid alignment that integrates small Gaussian pose errors; for 3D scans a beam likelihood (Gaussian plus uniform mixture approximated by triangular distributions) against a triangle-mesh model via ray tracing (Sec. 2.1-2.2) |
| 資料關聯 | no explicit correspondences; beams are ray-cast into the current grid (2D) or triangle mesh (3D), and max-range beams also contribute (Sec. 2.2) |
| 時間表示 | discrete poses |
| 去畸變 | 原文未報告 (indoor robots moved at 10 cm/s to obtain adequate 3D point density) |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | triangle mesh from neighbouring scan points, simplified into planar polygons by randomized region-growing plane fitting (delta 30 cm, epsilon 2.8, gamma 10 cm) and merging of coplanar neighbouring polygons; plane candidates found from line-angle histograms and plane sweeps at 5 cm then 1 cm (Sec. 3) |
| 先驗資訊 | none (built structures assumed to contain large flat surfaces) |
| 可輸出幾何 | compact 3D polygonal model: large planar polygons and quads plus residual triangles for non-planar regions (Table 2) |
| 計算需求 | Sieg Hall (1,933,018 points): 51 min 42 s plane extraction and 9 min 43 s polygon merging (Table 2); a naive plane extraction on 200,000 surfaces took over 10 h on a standard PC (Sec. 3.3) |
使用設備
原文使用的感測器、運算硬體與載具（equipment）。型號保留原文寫法，連結到設備頁中同一型號的歸併名稱；角色依原文用途分為方法輸入、資料集感測器、執行運算平台、參考或真值量測（reference or ground truth）與比較對象設備。
| 類別 | 型號（原文寫法） | 角色 | 資料集 | 原文規格 | 出處 |
|---|---|---|---|---|---|
| LiDAR | SICK PLS | 方法輸入 | 未標示 | measurement error below 20 cm; range resolution 1 cm | (Hähnel et al., 2003b, Sec. 4) |
| LiDAR | SICK LMS | 方法輸入 | 未標示 | measurement error below 5 cm; range resolution 1 cm | (Hähnel et al., 2003b, Sec. 4) |
| 載具平台 | Herbert | 方法輸入 | 未標示 | outdoor robot with one laser on a pan/tilt unit, angular resolution 0.25 deg | (Hähnel et al., 2003b, Sec. 2.3; Sec. 4; Fig. 4) |
作者報告的優勢與限制
優勢
- Reduction ratios between 100:14.2 and 100:4.1 over three data sets (Table 2).
- At the same polygon count as QSlim the model kept doors and planar walls that QSlim lost (Sec. 4; Fig. 10).
- Histogram-guided plane search gives more than an order of magnitude speed-up over naive plane extraction (Sec. 3.3).
- The probabilistic 3D scan matcher uses max-range beams and handles occlusion without special heuristics; the authors observed cases where ICP diverged but their matcher did not (Sec. 2.2; Sec. 5).
限制
- The model is limited to flat surfaces; measurements of non-flat objects are not corrected, so the model remains fairly complex (Sec. 6).
- No exploration strategy for full 3D acquisition (Sec. 6).
- Most of the mesh ruggedness comes from laser measurement noise, with some residual pose error (Sec. 3).
- (inference) Model fidelity is judged visually and by polygon counts; no comparison with surveyed geometry.
營建工程相關證據
未在營建工地驗證；資料來自 CMU Wean Hall、UW Sieg Hall 走廊與 Freiburg 校園約 40 m 乘 60 m 的建物外部。作者提到建築師與建物管理者可用 3D 模型做設計與使用研究（Sec. 1）。以平面擬合把雜訊網格簡化為牆、天花板與門的平面多邊形，是早期由行動雷射資料產生竣工建物模型的做法；ScienceDirect 的被引用清單也列出 Pătrăucean 等的竣工建模綜述（Pătrăucean et al., 2015）（推論）。
原文驗證環境：已完工建築
報告的性能數據
以下是原文作者報告的性能數值（author-reported results），不是本研究重新量測的結果。每張圖只並列同一個比較組（comparison group，同一張表、同一組實驗設定）內的方法；不同比較組之間的數值不可直接比較，也不構成排名。
本方法共出現在 1 個比較組，合計 18 筆紀錄。
Hähnel et al., 2003b · Table 2 本方法 18 筆
表格設定（擷取紀錄原文）：Statistics of the planar simplification for three data sets; times given as min:s in the paper and converted to seconds here; reduction ratio given as 100:x and stored as x (Hähnel et al., 2003b, Table 2)
Time (min) plane extraction (，CMU Wean Hall corridor (10 m traveled, SICK PLS) · Wean Hall
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Hähnel et al., 2003b 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| planar approximation and polygon merging (proposed)本方法原文提出 | 376 s原文指標寫法：Time (min) plane extraction (as written 6:16) | (Hähnel et al., 2003b, Table 2) |
來源
Hähnel et al., 2003b
（2003）Learning compact 3D models of indoor and outdoor environments with a mobile robotRobotics and Autonomous Systems (special issue: Best Papers of the Eurobot '01 Workshop), 44(1):15-27
DOI 10.1016/s0921-8890(03)00007-1
同儕審查已出版已讀全文經典查證後修正
相關版本
- earlier version：Workshop version in Proceedings of the Fourth European Workshop on Advanced Mobile Robots (EUROBOT'01), Lund, September 2001, as cited by Surmann et al. 2003 ref. [2] (not read) not_verified