Irma3D automated thermal 3D mapping
作者提出由機器人 Irma3D 全自動建立建築物熱影像三維模型的系統。平台以 Riegl VZ-400 地面雷射掃描儀為主感測器，上方裝 optris PI160 熱像儀與網路攝影機，以停走方式在各站掃描，每站再以 3DTK 的 6D SLAM 配準成同一座標系。熱像儀以燈泡陣列板做內參與相對掃描儀的外參校正，並以沿光線檢查遮擋的方式把溫度與顏色指派給點雲。站點選擇結合二維 NBV 探索與房間偵測後的三維體素 NBV 規劃，以減少天花板、地板與家具後方的遺漏；最後以行進立方體重建網格、映射溫度場並自動標出熱源。
本頁內容
Autonomous robot (Irma3D) that builds thermal 3D building models from stop-and-go Riegl VZ-400 scans registered by 3DTK 6D SLAM, with calibrated thermal and colour cameras, occlusion-aware projection, combined 2D and 3D next-best-view planning and marching-cubes reconstruction with heat-source detection.
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 | terrestrial 3D laser scanner (Riegl VZ-400)、thermal camera (optris PI160)、colour webcam (Logitech QuickCam Pro 9000)、2D laser scanner (SICK LMS100) for obstacle avoidance |
|---|---|
| 原文測試平台 | wheeled UGV (Irma3D on a Volksbot RT-3 chassis), stop-and-go scanning |
| 狀態估計 | Scan registration with 6D SLAM from 3DTK (The 3D Toolkit) for the scanner poses; GMapping under ROS for robot localization during exploration (Sec. 3.2.5, 4.1) |
| 資料關聯 | scan matching of stop-and-go 3D scans in 3DTK (details in cited work); calibration board detected in scans by RANSAC plane fitting plus ICP of a plane model (Algorithm 1) |
| 時間表示 | stop-and-go: static 360 deg scans, images taken during a return rotation after each scan |
| 去畸變 | 不適用 (static scans at each position) |
| 迴圈閉合 | as provided by 6D SLAM in 3DTK (not described in this paper) |
| 全域最佳化 | 6D SLAM registration in 3DTK (described in cited work) |
| 地圖表示 | registered 3D point cloud with reflectance, thermal and colour values; 0.2 m voxel model for 3D NBV planning; marching-cubes mesh with mapped temperature field (Sec. 4.3, 6) |
| 先驗資訊 | none about the building (exploration from a blank map); room-detection height chosen manually (2.5 m) |
| 可輸出幾何 | thermal and colour 3D point cloud of a building floor and reconstructed thermal surface model with automatically detected heat sources |
| 計算需求 | 原文未報告; reconstruction time 6.2 to 55.1 s depending on subdivision (Table 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 LMS100 | 方法輸入 | 未標示 | 2D laser scanner at the front for obstacle avoidance | (Borrmann et al., 2014, Sec. 3.1) |
| 地面雷射掃描儀（TLS） | Riegl VZ-400 | 方法輸入 | 未標示 | field of view 360 deg x 100 deg; stated accuracy 5 mm; head rotation used to carry the cameras | (Borrmann et al., 2014, Sec. 3.1, Sec. 6) |
| 相機 | Logitech QuickCam Pro 9000 | 方法輸入 | 未標示 | 1600 x 1200 video resolution; 10 images per camera per 360 deg | (Borrmann et al., 2014, Sec. 3.1) |
| 熱像儀 | optris PI160 | 方法輸入 | 未標示 | 160 x 120 px, thermal resolution 0.1 degC, 7.5 to 13 um, 120 Hz, accuracy 2 degC, FOV about 40 deg x 64 deg | (Borrmann et al., 2014, Sec. 3.1) |
| 載具平台 | Irma3D (Volksbot RT-3 chassis) | 方法輸入 | 未標示 | mobile robot carrying scanner and cameras | (Borrmann et al., 2014, Sec. 3.1; Fig. 2) |
| 其他 | calibration board with 30 lamps (12 V, 4 mm bulbs) and chessboard pattern | 方法輸入 | 未標示 | 500 mm x 570 mm board on a tripod for intrinsic and extrinsic thermal and colour camera calibration | (Borrmann et al., 2014, Sec. 3.2.1-3.2.2) |
作者報告的優勢與限制
優勢
- Complete autonomous pipeline from exploration and data acquisition to a thermal 3D model and heat-source detection (Sec. 7)
- 3D NBV planning left no unseen voxels visible from any candidate position in room 1, while 2D-only exploration left 703 (Table 2; Sec. 5)
- Room-based switching between 2D and 3D planning keeps memory and computation low (Sec. 4)
- Cited prior work reports positional error below 4 cm for the 6D SLAM registration even in large outdoor environments (Sec. 6)
限制
- Each scan takes 3 min 15 s, so the number of scanning positions must be minimized (Sec. 2.2, 5)
- Reconstruction precision is insufficient for furniture, monitors and people; outliers seen through windows reduce point density until removed by clustering (Sec. 6.3, 6.5)
- Low thermal-camera resolution (160 x 120) makes calibration inaccuracies matter and causes projection errors at edges (Sec. 3.2.4)
- Automatic interpretation of thermal flaws is still open; the work is fundamental research, not a ready product (Sec. 7)
營建工程相關證據
應用於既有建築的熱能檢測與節能改善，屬竣工建築的自動化掃描；三維幾何來自停走式地面雷射掃描與 6D SLAM 配準，論文本身未評估配準或幾何精度，只引用先前研究的定位誤差（Sec. 6）。
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 4 個比較組，合計 31 筆紀錄。
Borrmann et al., 2014 · Table 2 本方法 15 筆
表格設定（擷取紀錄原文）：Two additional exploration experiments in room 1 from the same start position, one with 2D exploration only and one with 2D plus 3D NBV planning (Borrmann et al., 2014, Table 2)
O (occupied voxels)，authors' Irma3D exploration data · room 1, Scan 1
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Borrmann et al., 2014 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Borrmann et al., 2014, Table 2)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| 3D NBV (combined 2D and 3D exploration)本方法原文提出 | 4068 voxels | (Borrmann et al., 2014, Table 2) |
| 2D NBV only本方法 | 4033 voxels | (Borrmann et al., 2014, Table 2) |
Borrmann et al., 2014 · Table 1 本方法 9 筆
表格設定（擷取紀錄原文）：Voxel counts of the 3D model of room 1 during the full exploration run (0.2 m voxels, field-of-view constraint of the thermal camera, stop threshold Vmin = 15 voxels) (Borrmann et al., 2014, Table 1)
O (occupied voxels)，authors' Irma3D exploration data · room 1, Scan 1
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 Borrmann et al., 2014 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 |
|---|---|---|
| combined 2D and 3D NBV exploration本方法原文提出 | 4091 voxels | (Borrmann et al., 2014, Table 1) |
Borrmann et al., 2014 · Table 3 本方法 6 筆
表格設定（擷取紀錄原文）：Marching-cubes reconstruction of part of the dataset with different spatial subdivisions (Borrmann et al., 2014, Table 3)
execution time of the reconstruction algorithm，authors' Irma3D data (part of the dataset) · subdivision 50
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 Borrmann et al., 2014 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 |
|---|---|---|
| probabilistic marching cubes reconstruction本方法原文提出 | 6.2 s | (Borrmann et al., 2014, Table 3) |
Borrmann et al., 2014 · Text Sec.5 本方法 1 筆
指標time per scan (3 min 15 s)
資料集與序列authors' Irma3D exploration data · per scanning position
表格設定（擷取紀錄原文）：Duration of one 3D scan with thermal and colour image acquisition at a scanning position (Borrmann et al., 2014, Text Sec.5)
time per scan (3 min 15 s)，authors' Irma3D exploration data · per scanning position
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 Borrmann et al., 2014 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 |
|---|---|---|
| Irma3D stop-and-go scanning本方法原文提出 | 195 s | (Borrmann et al., 2014, Sec. 5) |
來源
Borrmann et al., 2014
（2014）A mobile robot based system for fully automated thermal 3D mappingAdvanced Engineering Informatics, 28(4):425-440
同儕審查已出版已讀全文經典查證後修正