在施工中工地（active construction site）且有外部參考的研究較少，而且各有限制。(Blum et al., 2021, Sec. IV-C; Tables I-IV) 以全測站追蹤機器人上的稜鏡，評估靜止機器人相對建築平面圖模型的定位，但摘要所稱誤差至少降低 30% 的比較（Table II 至 IV）使用的模型偏差，是作者把網格上下兩側結構人為拉開 0.3 m 模擬的，只有 Table I 使用未加偏差的平面圖；(Zhang et al., 2024b, Sec. 4.2.1-4.2.2; Table 4) 評估 FAST-LIO2 點雲對 BIM 的配準時，參考轉換是以工程師量測的每層 3 至 4 個結構特徵點為同名點，在 CloudCompare 人工粗配準後再做點對網格精配準得到的，驗證的是配準，不是點雲本身的精度；(Křemen et al., 2025) 以 TLS 評估高速公路工程土方堆置場的背包式 SLAM 點雲。依本站分類，清單中標示為施工中工地的 16 筆研究裡，有獨立參考量測者 3 筆，只有部分或稀疏參考者 4 筆。此外，有兩篇研究以 ConSLAM 的施工中工地序列評估：(Stührenberg & Smarsly, 2025) 使用 ConSLAM 的真值軌跡與 TLS 掃描，(Vega-Torres et al., 2024) 則以最終 ICP 所用的同一張 TLS 地圖當評估參考；ConSLAM 真值軌跡本身的不確定度未量化（見資料集一節）。
(Feng et al., 2025, Sec. 5.1, 5.3; Table 4) 在施工中的醫院門診大樓，以預設參數比較十種開源 3D LiDAR SLAM；但全文未說明實地參考軌跡的來源，也未說明計算 APE 前的軌跡對齊方式，所以其公尺級的絕對位姿誤差（absolute pose error, APE）不能解讀為絕對精度。此研究列在下方資料集清單。
laboratory targets: 24 MDF colour boards (300 x 300 mm, smooth and 40-grit rough) and cast concrete targets (dolomite and granite mixes, three roughness levels)(Hawley & Gräbe, 2022, Sec. 3.1, 3.2, 4.2)
physical cylinder test object with pre-marked scale for the movement-rate test(Hsieh et al., 2023, Sec. 4.2.2, Figs. 6, 7)
spherical targets (diameter 0.145 m) and black-and-white targets(Křemen et al., 2024, Materials and methods; Measurement and processing of the reference dataset)
Leica GZT21 black-and-white targets and Leica GMP111 mini prism(Křemen et al., 2025, Measurement and processing of the reference dataset)
UAV photogrammetry DSM (nadir and oblique images; platform and camera not reported)(Sammartano & Spanò, 2018, Tables 10, 14)
close-range photogrammetry (SfM) model of the tower (camera not reported)(Sammartano & Spanò, 2018, Table 3; Sec. Tower (A))
spherical targets, diameter 0.14 m (4 GCPs)(Štroner et al., 2025, Sec. 2; Sec. 2.5)
flatness defect test-bed boards (3 boards)(Tang et al., 2011, Framework, Flatness Defect Detection Test Bed; Fig. 3)
原文未報告 (engineers' on-site measurement of 3-4 structural landmarks as GCPs)(Zhang et al., 2024b, Sec. 4.2.1)
設備名稱依紀錄原文列出、不翻譯，以免改變型號與描述；括號內為研究與原文位置。
對齊與參考方式會改變結論
同一批點雲換一種對齊方式，結論可能不同。(Fahle et al., 2022, Sec. 4.4) 報告以整體 ICP 對齊可能掩蓋 SLAM 漂移，改用入口處約束的對齊才顯示出累積誤差。(Štroner et al., 2025) 則只以坑道兩端的四顆球靶做剛體轉換，再分別計算整體偏差、ICP 後偏差、中段剖面的系統偏移與雜訊。
多項機器人研究以任務指標呈現成效，例如分割的精確率、召回率與 F1，或語意分割的平均交並比（mean intersection over union, mIoU），而不是點雲的幾何精度 (Kim et al., 2022a; Hu et al., 2023)。(Chung et al., 2025) 報告的覆蓋率是相對熟練人員人工掃描的體素數比值，最終點雲也來自機載 TLS，而不是 SLAM 地圖。
比較基準也會影響改善幅度的讀法。(Chen et al., 2025a, Abstract; Table 7) 的摘要稱位置漂移比「傳統 SLAM 方法」降低 71.77%，但這個比較對象是 VINS-Mono；相對 RTAB-Map（0.4587 m 降至 0.405 m）只降低約 12%。只引用前者，容易高估相對其他 SLAM 方法的改善幅度（推論）。
作者在 Boston Dynamics Spot 上加裝 Ouster OS0 光達與 RealSense D415，以 Kitware LiDAR SLAM 並用 BIM 結構元件點雲初始化平面特徵地圖，使機器人相對 BIM 定位。另建 0.1 m 體素的機率式 3D 地圖，逐體素記錄 BIM 元件是否被光達證實，並在 Nav2 成本地圖中對已證實結構、未證實 BIM 元件與暫時障礙物給予不同代價；再以粒子群最佳化的下一最佳視角規劃拍攝目標物，交由視覺語言模型判斷是否安裝。兩組各 20 次模擬的執行成功率為 97.5% 與 100%，實地單次試驗 6 個目標全數完成；論文未報告定位誤差。
作者的適用性主張
原文未報告
任務需求與來源
原文未報告來源：原文未報告
作者報告的結果
execution success 97.50% (dynamic simulation) and 100% (BIM-discrepancy simulation), 6 of 6 targets in the real trial (Table 1)
found the only viable path through unbuilt BIM walls in all 20 trials of Sim. Experiment 2 (Sec. 4.1)
ablation: Non-Semantic variant fell to 81.67% and 55.00% execution success, No NBV Retries to 80.00% and 76.67% (Table 2)
VLM precision near 100% in simulation and correct inference of a curtain-covered window in the real trial (Table 1; Sec. 4.1)
限制
原文未報告
no localization or map error metrics; evaluation is task-level only (Sec. 6)
single real-world trial; simulated experiments lack Spot's proprietary low-level controller (Sec. 3.7.1, Sec. 6)
manual approximate initial pose relative to BIM; no global localization because of structural symmetry (Sec. 3.4.1, Sec. 6)
requires BIM of LoD 300 or higher; tolerance to BIM discrepancies not systematically evaluated (Sec. 6)
robot arm partly occluded the forward LiDAR view in the real trial (Sec. 4.1)
single-floor 2D navigation only; no active search for misplaced objects (Sec. 4.1, Sec. 6)
simulated VLM accuracy depends on rendering fidelity (reception desk 10-15%); real VLM result rests on six images (Sec. 4.2.2, Sec. 6)
平台與感測器（原文）
legged (Boston Dynamics Spot)；simulation (Unreal Engine 5 digital twin)；3D LiDAR (Ouster OS0, 64 beams, 90 deg vertical and 360 deg horizontal FoV) for SLAM and mapping；RGB-D camera (Intel RealSense D415) for inspection images；Spot built-in RGB, five monochrome, six depth and one infrared cameras used by Spot's own obstacle avoidance
Rendered depths agree with valid MVS depths for 86.15%, 98.12% and 99.86% of pixels at thresholds 1.25, 1.25^2 and 1.25^3 over 210 viewpoints while filling MVS depth holes (Table 1)
Depth-aware masking lowers the mean point distance to the SfM+MVS mesh from 0.09 u to 0.083 u (7.7%) and the final training loss by about 10% (Geometric Accuracy Evaluation; training-loss text)
Total modeling time 1 h 05 min 33 s versus 1 h 44 min 09 s for MVS plus surface reconstruction, about 37% less (Table 3)
Fewer voids and less blurring and noise in dark segments than the MVS mesh; about 5% higher Brenner sharpness than unmasked 2DGS (Visual Quality Comparison)
限制
Geometric accuracy is agreement with the MVS mesh designated as reference (No. 1); models have no absolute scale (unit u), the registration method is not described, and no TLS, total station or other survey reference is used (Geometric Accuracy Evaluation)
2DGS training time (about 49 min) precludes real-time structural updating or inference (Limitations and Future Research Directions; Table 3)
Images give no useful supervision in complete darkness (Limitations and Future Research Directions)
Validated on a single shield tunnel scenario (Limitations and Future Research Directions)
Side-view renderings stay blurred without side-view images, and the Gaussian model alone is not suited to physical measurement (Shield Tunnel Gaussian Model)
In bright regions segment joints are rendered slightly less sharply than in the MVS mesh (Visual Quality Comparison)
Masking percentile p and TSDF voxel size are not reported; data and code available only on request (Methods; Data Availability Statement)
平台與感測器（原文）
UAV (DJI Mini 4 Pro) flown longitudinally near the tunnel central axis at 2.0 m/s；monocular camera (DJI Mini 4 Pro UAV camera; 1920x1080 video)
Hourly CV(RMSE) versus ground-truth (RENSA) simulations of 16.97, 8.80 and 6.58% (Building 1), 12.07, 16.20 and 15.49% (Building 2), 23.27, 11.64 and 1.79% (Building 3) for heating, cooling and electricity, all within the ASHRAE 30% limit (Sec. 4.2.1)
Post-GEMA P-SSIM of 0.8806, 0.9032 and 0.9003 (Table 1)
Annual energy within about 0.5 to 12.5% of the reference under 2x and 4x input downscaling; densification kept the 480x270 Building 1 model simulatable (Table 5; Sec. 4.2.3)
Densification takes 2.3 to 5.8 s on an RTX 3060 (Table 2)
限制
Largest dimensional-ratio deviation 20.47% (Rh/L, Building 1); footprint area error up to 12.81% and volume error up to 13.06% for single-detached houses; up to 13.97% area and 13.69% volume error for other typologies (Tables 3 and 4)
Reference geometry for Buildings 1 to 3 and all reference energy results come from the authors' RENSA models, not utility data (Sec. 1; Sec. 4.2.1); Buildings 4 to 6 are compared with 'ground truth image' dimensions of undescribed origin and have no energy reference (Table 4; Sec. 4.2.1)
Degrades in multi-object scenes and sparse flight coverage; reflectance-driven façade artifacts and holes; strong dependence on COLMAP initialization quality (Sec. 5)
Not evaluated on complex typologies such as high-rises or curved façades (Sec. 5)
P-SSIM ignores 3D spatial locality and can score visually poor meshes highly (Sec. 4.2.3)
工地機器人與機具定位construction robot localization and mapping
場域類型
模擬施工中工地原文未報告
平台與感測器
輪式地面機器人模擬光達（研究系統）
參考量測
實地 ATE 的軌跡真值來源未說明；地圖尺寸誤差以施工圖尺寸為參考，不是獨立量測（Sec. 4、5）。原文未報告
量測指標
幾何原文未報告任務原文未報告
場域與營建關聯實地測試在西安某醫院門診大樓的施工中工地，當時處於機電安裝與裝修階段，現場有工人、機具、鷹架、升降平台與推車，部分區域因地坪施工封閉；機器人以遙控方式行走 1004 m（2067 s）。模擬以同一棟大樓 CAD 圖建立 Gazebo 場景（標準層 292 m x 142 m x 6 m，行走 1293 m）。實地軌跡真值來源未說明；地圖尺寸以施工圖尺寸為參考，並非獨立量測（Sec. 4、Sec. 5）。
simulation ATE RMSE 5.40 m vs 12.33 m (LeGO-LOAM) and 20.27 m (F-LOAM) over 1293 m (Table 1)
real-site ATE RMSE 2.87 m vs 7.97 m (LeGO-LOAM) and 19.95 m (F-LOAM) over 1004 m (Table 5)
RMSE increased 10.00% with moving objects vs 19.95% (LeGO-LOAM) and 23.98% (F-LOAM) (Table 4)
map dimension error averaged 0.79% on site vs 2.33% (LeGO-LOAM) and 5.21% (F-LOAM) (Tables 6, 8)
runs on Jetson Xavier NX at 121.62 ms per frame, 7.20% slower than LeGO-LOAM (Table 8)
限制
原文未報告
dynamic objects still degrade accuracy; semantic segmentation and object tracking left for future work (Sec. 4.2.4, Sec. 6)
simulation omits dust, lighting and reflectivity effects and cannot reproduce Ackermann turning and stop-and-go motion (Sec. 4)
real-site run limited to 1004 m by safety closures; teleoperated rather than autonomous (Sec. 5.1, Sec. 5.2.5)
higher computational load: 121.62 ms per frame vs 65.63 ms for F-LOAM (Table 8)
residual z-axis error remains after loop closure (Sec. 5.2.1)
no comparison with tightly coupled LiDAR-inertial methods; loop-closure false positive and negative rates not quantified (Sec. 4.2, Sec. 6)
(inference) real-site ground-truth source not described, and the proposed-method RMSE values in Tables 1, 4, 5 and 7 are inconsistent with their own mean and STD
平台與感測器（原文）
wheeled UGV (Ackermann-steered construction robot base, teleoperated at 0.50 m/s on site)；simulation (Gazebo, robot moved at 1.00 m/s)；3D LiDAR RS-Helios-16P (16 beams, 10 Hz, +-15 deg vertical FOV, +-2 cm ranging)；nine-axis IMU at 200 Hz (recorded; the evaluated system is LiDAR-only)；camera (recorded; model not reported; not used by the method)
synthetic: R2 0.963 and RMSE 0.009 m vs best M3C2 (0.05 to 0.55 m scale) 0.955 and 0.010 m; 187 s vs 995 s (Table 4)
curved field target MAE 0.032 m vs 0.034 m for M3C2 (Sec. 4.4)
robust to 20% point removal (R2 0.956, RMSE 0.010 m) and to noise at 10% of deformation (R2 0.938, RMSE 0.012 m) (Table 5)
noise floor in stable regions: sigma 0.0035 m short-term field, 0.0151 m over two months (Table 6)
code and datasets shared on GitHub (Data availability)
限制
原文未報告
evaluation relies on a limited set of multi-temporal scans from one mine (Sec. 4.8)
ICP-based registration contributes systemic noise; real-scenario detection threshold set at 3.0 cm (Sec. 4.7, 4.8)
tunnel treated as a single entity without semantic segmentation (Sec. 4.8)
for tunnels longer than about 60 m the authors recommend splitting the cloud into sections (sliding window) because a single global volume causes memory and time bottlenecks (Sec. 4.3)
planar field target shows no gain over M3C2 (both MAE 0.022 m) (Sec. 4.4)
real-deformation case checked only qualitatively plus one manual point-to-point measurement (Sec. 4.5)
at noise 20% of deformation R2 drops to 0.858 and RMSE rises to 0.018 m (Table 5)
平台與感測器（原文）
handheld；handheld Hovermap ST LiDAR SLAM scanner (Table 1: FoV 360 x 290 deg, range 0.40 to 100 m, LiDAR accuracy +/-30 mm, mapping accuracy +/-15 mm in typical underground and indoor environments, SLAM drift +/-0.03%, up to 300,000 pts/s single return and 600,000 pts/s dual return)
perceived usefulness 61%, ease of use 68%, adoption potential 86% among 36 construction management students (Abstract)
authors claim lower computational load than other state-of-the-art LIO systems (Abstract)
authors claim greater versatility and usability than stationary and drone-based LiDAR platforms in complex indoor environments, at significantly lower implementation cost (Abstract)
限制
摘要未提及
concerns about cost and workflow integration noted by participants (Abstract)
(inference) no geometric accuracy evidence in abstract
巡檢與缺陷檢測robot-aided visual defect inspection of civil infrastructure (defect localization and sizing on a bridge and a parking garage)
場域類型
基礎設施含任務層驗證2 structures (1 bridge, 1 parking garage), UGV data
平台與感測器
輪式地面機器人光達（研究系統）相機IMU
參考量測
只以 7 個明確特徵的長度對照現場量測（儀器未說明）；沒有軌跡、地圖或缺陷尺寸的真值（Sec. 5.5）。7 on-site length measurements of unambiguous features (instrument not stated); no trajectory, map or defect ground truth
量測指標
幾何reference-free point-to-plane surface thickness on extracted planes and CloudCompare density (Table 7)任務defect hull areas, bounding boxes and crack lengths (Tables 5-6, no ground truth); 7 proxy feature lengths versus on-site measurements (Table 9)
場域與營建關聯目標期刊中把 LiDAR-相機-慣性 SLAM 用於營運中民用基礎設施巡檢的研究：實測對象為加拿大 Kitchener 的一座混凝土箱梁鐵路橋與一座停車場（非施工中工地），另公開 8 處、18 筆紀錄的資料集。可補足營運中基礎設施巡檢的證據；在該補缺批次之前，同一證據叢集的已核實紀錄中屬基礎設施者僅（Hawley & Gräbe, 2022），其為鐵路隧道研究。作者相關博士論文（UWSpace，題名 Towards SLAM-Centric Inspection of Infrastructure，與本文不同）不作為本文證據。
作者主張、報告結果與限制
研究重點
作者提出以 SLAM 為中心的機器人輔助目視巡檢流程：線上 LiDAR-相機-慣性 SLAM（重新實作 LVI-SAM 架構，作者不主張其新穎性）、離線批次軌跡精修（以歐氏距離與 Scan Context 偵測迴圈）、與 SLAM 地圖解耦的巡檢點雲生成、影像缺陷分割（DIS-YOLO 與 SAM），以及把像素以加速射線追蹤投影到無序 LiDAR 點雲而不需建網格。實測對象為加拿大 Kitchener 的 Park Street 混凝土箱梁鐵路橋與 Duke Street 停車場，資料以 UGV 蒐集。作者明言缺乏軌跡真值，未評估漂移或 APE；地圖品質以自訂的平面點到平面厚度指標評估，尺寸精度以 7 個明確特徵對照現場量測，平均絕對誤差 3 cm（2.7%）；缺陷本身的尺寸沒有真值。
作者的適用性主張
accurate defect localization, dimensional quantification and dense inspection maps in real-world scenarios (abstract)
任務需求與來源
原文未報告 as an acceptance requirement; GSD analysis targets minimum detectable crack widths of 2.1 mm (garage) and 3.5 mm (bridge) (Table 3)來源：原文未報告
作者報告的結果
Full offline refinement reduced the point-to-plane surface thickness from 4.164 to 3.647 cm (garage) and from 6.175 to 4.231 cm (bridge), with CloudCompare density rising accordingly (Table 7)
Seven proxy measurements of unambiguous features differed from on-site measurements by 3 cm (2.7%) on average, SD 2.3 cm (Sec. 5.5.3, Table 9)
Hardware time synchronization via microcontroller and documented sensor selection; base module about 958 g (Sec. 4.2-4.3)
Public dataset of 18 recordings, 3294 s, 327.12 Gb and about 84k images; the text says 8 locations (4 labs, 2 bridges, 1 garage, 1 open outdoor site) but Table 4 lists 7 location rows (Sec. 5.2, Table 4)
Authors report an average ray-tracing speed-up of 9000x over naive ray-tracing (Sec. 5.5.2); per-image ratios computable from Table 8 are about 390x, 1820x and 7040x (mean about 3080x), so the 9000x figure cannot be reproduced from Table 8 (inference from table values)
限制
no trajectory ground truth; proxy features instead of defects; planar-defect assumption; UGV only in evaluation
No ground truth for trajectories or defects; drift and absolute pose error were not evaluated (Sec. 5.5, Sec. 6)
Sec. 3 states the work relies on qualitative assessment to validate the approach because no benchmarks exist for inspection maps
Map quality is judged with a self-defined, reference-free point-to-plane metric on extracted planes (Sec. 5.5.1)
Measurement accuracy is validated on 7 proxy features, not on the defects; true in-field defect errors may be higher (Sec. 5.5.3)
Defect area assumes planar defects; defect-tracking accuracy after map alignment is not assessed (Sec. 3.3.2)
Reliable SLAM remains a challenge and the engineering burden is high (Sec. 6)
Table 9 row 3 is internally inconsistent (site 2.41 m and map 2.68 m, but abs error 0.08 m and 3.3%)
(inference) Both evaluated datasets used the UGV only, so the platform-agnostic claim rests on design arguments rather than cross-platform evaluation
Sec. 4.1 text swaps the camera descriptions relative to Table 1 (12.6 MP with 185 x 140 deg in the text vs Table 1: 12.3 MP inspection camera with a 50.8 x 38.6 deg lens); Table 1 is taken as authoritative
Sec. 5.5.3 links the Table 5 defect to measurements 4 and 5, but the matching dimensions correspond to measurements 2 and 3
平台與感測器（原文）
UGV (Clearpath Husky) for both evaluated datasets (Sec. 5.1)；handheld rig with a backpack for computer and battery (Sec. 4.2); handheld recordings exist in the public dataset (Conestogo bridge, Structures lab) but are not evaluated (Table 4)；USV written as 'Huron from Clearpath Robotics' (Sec. 4.2); no USV recordings are listed in Table 4；3D LiDAR Velodyne VLP-16 Lite (UGV adds an upward VLP-16)；SLAM camera FLIR Blackfly S 3.2 MP with 185 x 140 deg Fujinon lens；inspection camera FLIR Blackfly S 12.3 MP with 8 mm lens；IMU Xsens MTi-30 (400 Hz)；UGV adds an upward high-resolution RGB camera and a FLIR ADK infrared camera；Teensy 3.6 microcontroller for hardware time synchronization
場域與營建關聯於既有建物（現代建築遺產）以商用行動式 SLAM 掃描器與手機影像實測比較，支持「高斯表示作為量測點雲之上的視覺化層，而非取代」的論點；但研究未量化任一輸出的幾何誤差，SLAM 掃描的軌跡、閉環與精度也都未報告，因此不能作為 SLAM 點雲或 3DGS 幾何精度的證據，只能支持視覺化與 VR 效能面向的比較。
Polycam 3DGS Laplacian variance 2021.88 versus 1181.53 for the GeoSLAM point cloud (HBIM 3493.74) (Sec. 4.2)
In Unity VR, 3DGS loads in 1 s versus 40 s, runs at 60-90 FPS versus 15-60, uses 13.6 MB versus 562 MB memory, 12-25% versus 35-55% GPU overhead, about 10 ms versus about 35 ms latency (Table 3)
LiDAR point cloud with SAM gives higher segmentation recall against HBIM than 3DGS with GMM (Sec. 4.3, Fig. 10)
Smartphone-based 3DGS lowers acquisition cost and technical barrier (Sec. 4.1)
限制
3DGS showed reduced geometric precision in complex or occluded areas and lower segmentation recall than LiDAR (Sec. 4.3, Sec. 6)
3DGS depends on well-calibrated, high-resolution images with consistent lighting and full coverage; exposure variation and occlusion may cause local errors (Sec. 5.3)
High GPU demand of converting dense LiDAR clouds to Gaussians forced workaround pipelines with resolution trade-offs (Sec. 5.3)
3DGS lacks standard interoperability with HBIM, GIS and semantic heritage datasets (Sec. 5.3)
LiDAR point clouds in VR showed fragmentation, transparency artifacts and slower response (Sec. 4.4)
平台與感測器（原文）
mobile scanner (GeoSLAM HORIZON RT), carrying mode not reported；smartphone multi-view photography (iPhone 12 Pro)；GeoSLAM HORIZON RT mobile SLAM scanner with integrated camera for panoramic images (LiDAR workflow; footnote 12 links the FARO GeoSLAM ZEB Horizon RT page)；iPhone 12 Pro smartphone camera, 124 images at 3024 x 4032 px (3DGS workflow)
場域與營建關聯礦業坑道（地下工程類環境），以高規格測量控制作獨立參考；可作隧道與地下空間 SLAM 掃描精度的代表證據，但非施工中隧道，作者也提醒牆面平滑的建築物結果可能不同（Sec. 4）。
作者主張、報告結果與限制
研究重點
作者在捷克 URC Josef 地下研究中心約 120 m 的不規則岩壁坑道，以 Leica MS60 全測站建立控制網並以 Leica P40 建立參考點雲（以全測站量測 24 個檢核點驗證，其中 1 點經目視檢查剔除，RMSD 1.4 mm），比較四款商用 SLAM 掃描儀與兩款靜態掃描儀。所有點雲僅以兩端四顆球靶做剛體轉換，再計算整體 RMSD、ICP 後 RMSD、中段 3.5 m 區段的系統偏移、雜訊與平滑後誤差。新世代 SLAM 掃描儀整體偏差與靜態掃描同級甚至更好（12 至 24 mm 對 22 至 27 mm），但雜訊較高、需平滑，誤差在兩端控制點中間最大，呈「香蕉形」變形；SLAM 掃描約 2 分鐘，靜態掃描約 2 小時。
作者的適用性主張
authors state new-generation SLAM scanners are highly suitable for enclosed spaces
任務需求與來源
原文未報告 (no project tolerance); cites sub-cm to several-cm precision needs generally來源：原文未報告
作者報告的結果
RMSD after GCP transformation 12-24 mm for SLAM (excluding the erroneous Geoslam run) vs 22-27 mm for static scanners (Abstract; Table 3)
Hovermap two-way: RMSD_PR 12 mm, RMSD_CP-XYZ 7 mm, MaxdXYZ 17 mm (Table 3)
SLAM gives more complete coverage of rugged hollows than stop-and-go static scanning (Sec. 3.3, Fig. 5)
SLAM capture about 2 min for the 120 m trail vs about 2 h for each static scanner (Sec. 2.4, Sec. 4)
after simple MLS smoothing, local accuracy of SLAM clouds approaches that of static scanners (Sec. 3.4, Table 3)
限制
single site; distributor-operated SLAM runs; mining not construction
SLAM clouds noisier; smoothing improves accuracy but may remove sharp edges and cannot recover features at or below the noise length scale (Sec. 4, Sec. 5)
one of three Geoslam runs obviously erroneous (~149 mm transverse deviation) (Sec. 3.4)
largest deviations mid-way between end GCPs (banana-like deformation) (Sec. 3.4, Fig. 6)
SLAM data collected by distributors' technicians; single 120 m test site (Sec. 2.2)
SLAM scanners give no in-method check; authors recommend at least two passes or extra checkpoints (Sec. 4)
findings limited to rugged mine walls; smooth-walled buildings may behave differently; three runs per SLAM scanner do not allow rigorous statistics (Sec. 4)
NavVis clouds are smoothed by the vendor software, flattering its noise profile (Sec. 3.3, Fig. 4)
平台與感測器（原文）
handheld；wearable；commercial SLAM scanners: GeoSLAM ZEB Horizon RT (Velodyne VLP-16, 16 channels), NavVis VLX 2 (two 16-channel sensors, 4 cameras), Emesent Hovermap ST-X (32 channels), FARO Orbis (32 channels, 360-degree camera)；static TLS: Trimble X7, FARO Focus Premium 70；reference: Leica ScanStation P40 + Leica Nova MS60 total station; four 0.14 m spherical GCP targets; Leica GZT21 black-and-white targets
幾何mean Chamfer distance 0.119 m (SD 0.136 m) vs TLS; position drift 0.405 m over 92.77 m (71.77% below VINS-Mono, ~12% below RTAB-Map)任務collision-free navigation; mAP 73.7% (2D) and 62.9% (3D)
authors claim suitability for indoor inspection; early-stage construction features (e.g., scaffolds) mentioned as assumption
任務需求與來源
原文未報告來源：原文未報告
作者報告的結果
position drift 0.405 m (0.44%) over 92.77 m, 71.77% lower than VINS-Mono baseline and ~12% lower than RTAB-Map (0.4587 m) (Table 7)
reconstruction mean Chamfer distance 0.119 m (SD 0.136 m) to TLS, 42.83% lower than RTAB-Map (reconstruction accuracy table)
collision-free navigation over a 92.77 m trip in 195 s (Table 6)
限制
completed occupied building; single corridor
battery: about 20% (1000 mAh) per inspection trip, at most five trips per charge (Sec. 4.2.1)
assumes limited pitch, roll and z-displacement; scan rectification does not correct yaw drift (Sec. 3.2, Sec. 3.2.1)
step-wise navigation needed to avoid local optima or navigation failure (Sec. 4.2)
ceiling areas not reconstructed because of RGB-D range and illumination; point density lower than RTAB-Map after noise removal (Sec. 4.2.3)
path planning not optimized for legged motion; one small quadruped with custom sensors limits generalizability; map shifting expected in featureless spaces (Sec. 5)
(inference, not stated by the authors as a limitation) localization reference derived from loop closure and scale alignment to the TLS cloud, so it is not fully independent of the evaluated data (Sec. 4.1)
機器人資料蒐集與掃描規劃scaffold point cloud acquisition (scan planning and execution)
場域類型
受控實驗施工中工地含任務層驗證1 controlled outdoor site (Site F) for termination-criteria and manual-comparison trials + 1 large-scale construction site; 6 sites for detector training/testing data
平台與感測器
四足機器人光達（研究系統）IMURGB-D 或深度相機地面光達（TLS）
參考量測
以熟練人員的人工掃描為比較基準計算覆蓋率，不是幾何精度參考。comparison with manual scanning by skilled workers
量測指標
幾何coverage rate (relative to manual scans)任務number of scan positions; coverage 106.1% and 96.8%
場域與營建關聯Site F 由作者描述為小型施工工地（Sec. 4.3.1）；Sec. 4.3.2 在同一場址改動部分鷹架位置並加入類似鷹架的障礙物後重複試驗，屬受控條件。另於一處大型施工工地測試（Sec. 4.3.3），作者致謝 Ssangyong Construction 提供戶外實驗工地。覆蓋率為相對人工掃描平均體素數的比值（0.05 m 體素，Eq. 8），不是絕對幾何精度；大型工地覆蓋率在摘要與內文寫 96.8%，Table 7 為 96.6%。
作者主張、報告結果與限制
研究重點
ASPAR 以四足機器人先自主探索工地並以 LIO-SAM 建立 3D SLAM 地圖，再把地圖投影為鳥瞰圖以 YOLOv8-OBB 即時偵測鷹架單元。系統依多層網格與貪婪法選定掃描站位並以 A*、旅行推銷員問題規劃路徑，最後在各站以機載 TLS 進行靜態掃描。此設計將 SLAM 用於規劃與導航，而最終幾何成果仍採用 TLS。
作者的適用性主張
authors claim real-world applicability on construction sites
任務需求與來源
coverage relative to manual TLS scanning by skilled workers (voxel count ratio, 0.05 m voxels, Eq. 8)來源：原文未報告
作者報告的結果
scaffold detector on BEV images: precision 0.985, recall 0.955, F1 0.971, [email protected] 0.985 on 100 site-F test images (Table 3)
all six scaffolds registered in five exploration runs; average precision 0.971, recall 1, F1 0.985 (Table 4)
average coverage 106.1 % of the manual mean with 5.4 vs 4.7 scan positions (Table 6)
large-scale site: coverage 96.6 % (Table 7; 96.8 % in text) with one extra scan position
detector trained on Ouster OS0-128 data worked on Velodyne VLP-32 field data (Sec. 4.3.1)
handles moving-object noise by refreshing the planning map with the latest raw scans (Sec. 3.2.1)
限制
geometric accuracy of SLAM map not the output; details not read
small or dynamic obstacles may block planned paths, requiring re-planning (Sec. 3.4)
TLS line of sight can be blocked by the mobile LiDAR, requiring robot rotation (Sec. 3.4)
obstacle detection based on height and density cannot detect non-obstacle hazards such as falling hazards or ground conditions (Sec. 4.4)
noise from movable obstacles in the SLAM map affects scan-position evaluation (Sec. 4.4)
repeated exploration gives slightly different SLAM maps and hence different scan plans (Sec. 4.4)
(from Table 6 and Table 7 values; not stated by the authors as a limitation) automated scanning took longer than manual scanning: 720 s vs 592 s on average, 1079 s vs 549 s at the large site
平台與感測器（原文）
legged (Unitree Aliengo)；3D LiDAR (Velodyne VLP-32)；IMU (Microstrain 3DM-GX5-AHRS)；IR depth camera (Azure Kinect)；TLS (FARO Focus M70) on robot；3D LiDAR (Ouster OS0-128), detector training and test data only
CatBoost multi-output regression with five-fold cross-validation: the text states R2 above 0.998 and RMAE below 0.01 for location and scale parameters of sphericity and density (Sec. 3.4); Table 7 column labels appear swapped and, read that way, the density location RMAE is 0.0114, so the claim is not fully consistent with Table 7
speed identified as the most influential factor; very slow (2 km/h) and very fast (8 km/h) walking degraded quality (Sec. 3.3, 5)
quantifiable basis for optimizing acquisition strategy (Abstract, Sec. 5)
限制
原文未報告
all data from a single metro tunnel system, limiting generalizability (Sec. 4.3)
limited number of experimental samples (Sec. 5)
static regression without SLAM temporal dynamics or error propagation (Sec. 5)
only geometric features, no colour or semantics (Sec. 5)
hyperparameters chosen without extensive tuning (Sec. 3.4)
平台與感測器（原文）
backpack carried by one operator (same operator throughout), walking about 2 to 8 km/h on flat track or between rails；OmniSLAM R6 commercial backpack SLAM system: rotating 32-beam LiDAR (maximum range 120-300 m, up to 640,000 points/s, up to 10,000 points/m2), inertial navigation fused by LIO, RTK-SLAM positioning module, 360 deg panoramic imaging unit with two 1-inch CMOS sensors; stated absolute accuracy 3 cm, relative better than 1 cm
SLAM scanning is a more practical alternative to static TLS for earthwork monitoring
任務需求與來源
原文未報告 (RTK accuracy 10 mm horizontal / 40 mm vertical cited as context)來源：原文未報告
作者報告的結果
GCP-based RMSD ~17 mm close to declared 15 mm mapping accuracy (Abstract)
operationally simpler and faster than static TLS for earthwork monitoring (Abstract)
限制
single site, single device; no volume validation read
systematic height shifts in RTK datasets (Abstract; Conclusions)
smoothing gave only slight improvement on rough soil surface (Results)
manufacturer's 15 mm mapping accuracy not achieved; the heap top was reached by a single narrow path, leaving top and slopes weakly connected (mean top vs sides height offset -22 vs -3 mm for RTK_1, -8 vs -1 mm for GCP runs) (Discussion)
only two GCPs placed on the heap top, not optimally; walking along the top edge was not feasible for operator safety (Discussion)
平台與感測器（原文）
backpack；Emesent Hovermap ST-X (32-channel LiDAR) on a backpack；Trimble R12i GNSS-RTK receiver on the backpack, CZEPOS network RTK (RTK runs only)；reference: Leica ScanStation P40 TLS (12 stations) and Leica Nova TS60 total station
ConSLAM translational APE RMSE 10.21-15.68 cm across sequences 2-5 vs ConSLAM ground truth (Table 5)
against SLAM2REF ground truth, 5.97 cm (seq 2) and 12.57 cm vs 18.02 cm for LIO-SAM (seq 5) (Table 6)
point-cloud inlier RMSE (0.3 m threshold) 6.3-7.9 cm vs TLS; office 6.57 cm vs 8.42 cm for LIO-SAM (Table 7)
a wrongly placed wall in the office BIM did not corrupt the map because low-compliance matches were rejected (Sec. 6, Fig. 16)
authors state the accuracy meets the +-10 cm needed for half-cell potential corrosion surveys except for a few outliers (Sec. 6)
open-source code (GitHub)
限制
原文未報告
not uniformly better than LIO-SAM: LIO-SAM had lower APE RMSE on ConSLAM sequences 3 and 4 against both ground truths (Tables 5-6) and lower inlier RMSE on sequences 3 and 4 (Table 7)
error grows from 10.21 cm (seq 2) to 15.68 cm (seq 5, 4.5 months later) as the site departs from the BIM (Sec. 6)
AprilTags must be placed manually at matching locations in the building and the BIM; BIM matching starts only after a tag is seen (Sec. 3.2, Sec. 7)
parameters chosen experimentally; lidar reflections in windows create fictitious walls (Sec. 4.4, Sec. 6, Sec. 7, Fig. 15)
BIM scan matching skips 23.6-33.1% of keyframes on ConSLAM because of processing time (Table 8)
(inference) the ConSLAM BIM was derived from the TLS scan of sequence 2, so it is an as-built model and likely favours sequence 2
平台與感測器（原文）
legged (Unitree A1-based 'IDOG')；handheld (ConSLAM dataset)；3D LiDAR (Velodyne VLP-16)；9-DoF IMU (LORD MicroStrain 3DM-GX5-25; Xsens MTi-610 in ConSLAM)；camera for AprilTag detection (Intel RealSense D435i; Alvium U-319c 3.2 MP in ConSLAM)；reference: Faro Focus S 70 TLS (office); Leica RTC 360 TLS scans of ConSLAM
Per-class mean wRMSE of 0.97% (wall), 0.85% (pillar) and 2.21% (beam), mean 1.34%, in the completed scene_1; these are the values behind the abstract's '1% to 2.2%' (Table 5, Sec. 4.3.2)
Mean segmentation AE of 8.7 IoU points in scene_1 and mean progress-tracking AE of 5.7 in the ongoing scene_2 (Tables 5, 7)
BIM-generated ground-truth masks remove manual labelling for progress tracking (Sec. 3.3.2, Sec. 4.3.1)
NeRF renders tall walls and columns more completely than iPhone flash LiDAR or photogrammetry point clouds (qualitative, Sec. 4.2.2, Fig. 19)
限制
NeRF rendering artefacts with moving objects or complex lighting affect evaluation; data were captured after work hours to avoid workers (Sec. 5, Sec. 6)
Inaccurate NeRF-BIM synchronization degrades IoU-based tracking and measurement; a 2 cm shift changes IoU by 0.03 (pillar, x or y) or 0.05 (wall, z), so about +-6 cm (x, y) or +-4 cm (z) keeps worst-case tracking error within 10% (Sec. 4.3.2, Sec. 6)
Alignment relies on checkerboards installed on site and replicated in BIM (Sec. 3.2.1, Sec. 6)
BIM holds only the final poured state, so formwork and rebar scenes were evaluated mainly by element counting; segmentation accuracy of incomplete elements needed manual labels (Sec. 4.2.1, Sec. 4.3.1)
Beams are less visible in the videos and show the largest errors (mAE 14.6, mwRMSE 2.21%) (Sec. 4.3.2, Table 5)
Rendering quality drops for the large outdoor UAV scene (site3_drone PSNR 19.27, SSIM 0.509) (Table 3)
平台與感測器（原文）
handheld；UAV；smartphone camera video (iPhone 15 Pro, 1080 x 1920)；UAV camera video (DJI Mini 2 Pro as written, 1920 x 1080)；iPhone 15 Pro flash LiDAR with Pix4D apps (qualitative comparison only)
場域與營建關聯測試場為 144 m² 室內空間（含牆、柱、外牆面、家具與走動人員），並非施工中工地；作者認為可用於施工進度監測與設施管理，但未在工地驗證（Sec. 4.1）。
作者主張、報告結果與限制
研究重點
作者以四足機器人搭載 Ouster OS1-128 3D 光達與一具 2D 光達，採「解耦」配置：2D 光達以 ROS2 Gmapping 即時建立占據格網地圖供導航與避碰，3D 光達資料則在掃描後以 Lidarslam_ros2 離線建圖（掃描匹配前端搭配 IMU 預積分與失真校正，後端為具迴路偵測的位姿圖最佳化）。掃描規劃將 .stl 模型體素化後隨機產生視點，以射線投射處理部分遮蔽，再以貪婪法挑選視點並以 TSP 求最短路徑。在 144 m² 室內測試場選出 3 個視點、理論覆蓋率 95%，實際掃描 8 分鐘。以雷射測距儀為真值，5 組距離的平均偏差為 0.041 m，引自前期研究的 TLS 為 0.028 m；152 mm 網格覆蓋率 83.12%（TLS 85.35%），但 25 mm 與 13 mm 網格的覆蓋率遠低於 TLS，作者也指出點密度低於 TLS。耦合與解耦兩種策略僅作定性比較。
作者的適用性主張
原文未報告
任務需求與來源
原文未報告來源：原文未報告
作者報告的結果
mean discrepancy 0.041 m to laser rangefinder over 5 distances vs 0.028 m for TLS from Zhai et al. 2024 (Table 2, Sec. 4.2)
152 mm grid coverage 83.12% vs TLS 85.35%; slabs 84.32% and 95.82% vs TLS 75.11% and 85.34%; Wall 1C captured where TLS had NIL (Table 3)
8 min automated scan plus 20 to 30 min mapping vs about 30 min TLS field work plus 30 min registration (Sec. 4.2)
3 viewpoints giving 95% theoretical coverage of the 144 m2 test site (Sec. 4.1)
all listed robotic densities exceed the 18.84 x10^-4 /mm2 minimum (Table 4)
限制
原文未報告
robot deviates from planned scan positions because of IMU precision, coordinate conversion between .stl and Gmapping frames and collision avoidance (Sec. 5)
onboard system-on-module has limited memory and compute for large scenes (Sec. 4.2, 5)
coverage at 25 mm and 13 mm grids (12.20% and 3.60%) far below TLS (74.53% and 67.61%) (Table 3); point density lower than TLS, which the authors attribute to an algorithmic limitation that keeps the Ouster OS1 hardware from being fully used (Sec. 4.4, Table 4)
(inference) TLS baseline comes from a different study (Zhai et al. 2024), not collected in the same campaign
(inference) coupled vs decoupled strategies are compared only qualitatively (Table 1)
平台與感測器（原文）
legged (quadruped robot; model not reported)；3D LiDAR Ouster OS1-128-Rev-07 (360 deg horizontal x 45 deg vertical FOV), mounted at 0.5 m, used for 3D reconstruction；2D LiDAR for navigation mapping (text gives 'e.g., RPLIDAR'; exact model not reported)；odometry and IMU data used by Lidarslam_ros2 (IMU source not specified; Sec. 5 mentions the quadruped robot's IMU)
作者以搭載 Ouster OS1-128 的 Boston Dynamics Spot，在走廊（44 m 開放路線）、地下室（49 m 封閉迴圈）與連接兩層的樓梯間（32 m）比較 KISS-ICP、FAST-LIO2 與 LIO-SAM-6AXIS。以 Leica TS30 全測站追蹤光達上方稜鏡取得參考軌跡，並提出時間同步與中斷資料拼接方法；另以 RIEGL VZ-400i 與 VZ-600i 地面掃描作為點雲參考。FAST-LIO2 在三種場景中有兩種 APE 最低、三種點雲平均距離皆最低（3.7 至 5.6 cm），點雲也最密；地下室的 RMS APE 為 9 至 35 cm，機器人上下樓梯時誤差明顯增加。
作者的適用性主張
authors state Spot + Ouster + open-source SLAM can automate geometric as-built documentation cost-effectively and may be used on construction sites
任務需求與來源
原文未報告來源：原文未報告
作者報告的結果
FAST-LIO2 lowest mean C2C in all three environments (3.7-5.6 cm) and lowest RMS APE in 2 of 3 (Tables 2-3)
method to synchronize total-station ground truth with SLAM data when line of sight is interrupted (Discussion)
FAST-LIO2 clouds were the densest (6.8-12.1 million points vs 0.27-0.36 million for LIO-SAM-6AXIS) with minimal noise (Table 3, Results)
FAST-LIO2 showed little drift when revisiting the cellar start despite no explicit loop closure, attributed to its large local map (Results, Discussion)
限制
single building; C2C after ICP alignment; Autowalk-defined routes
cellar RMS APE 9-35 cm; all algorithms struggled after the first curve (Results)
errors increase on stairs, partly from prism shifting with robot tilt (about 3-5 cm) (Results)
point clouds aligned to TLS by point pairs + ICP before C2C, which evaluates global quality but can absorb drift (Discussion)
KISS-ICP double surfaces near the cellar start; KISS-ICP and LIO-SAM-6AXIS strong shearing on the lower stairwell floor (Results)
FAST-LIO2 and KISS-ICP slightly poorer coverage at the hallway start, possibly due to the LiDAR minimum range (Results)
SLAM parameters (voxel and leaf sizes) tuned empirically for indoor density; trajectory alignment uses only the first 1.5 m (Methods)
(inference) single building, single run per environment
平台與感測器（原文）
legged (Boston Dynamics Spot, Autowalk missions)；3D LiDAR (Ouster OS1-128) with internal 6-axis IMU；reference: Leica TS30 total station with 360-degree prism；reference: Riegl VZ-400i and VZ-600i TLS
場域與營建關聯一般測繪儀器評估，於既有學術建築（美國西點軍校 Washington Hall 五、六樓）室內進行，未涉及施工中工地。行走距離越長誤差越大的結果，對大範圍室內竣工量測的控制點與路線規劃有參考價值（推論）。
作者主張、報告結果與限制
研究重點
作者在美國西點軍校 Washington Hall 學術建築評估 Leica BLK2GO 手持視覺加光達 SLAM 掃描儀。靜態測試以 Trimble SX10 量得掃描儀原點至約 2 m 與 15 m 標靶的距離，連續八次各五分鐘收集，RMSE 分別為 0.001 m 與 0.021 m，且未見隨機身溫度上升而變化的測距漂移。移動測試以五、六樓牆面標靶間距離與 SX10 比較，九種路線條件（開放、單圈、雙圈；單層或兩層；視覺加光達或僅光達）的 RMSE 為 0.033 至 0.128 m，並與累積行走距離呈現可能顯著的線性相關（R² 0.63 至 0.80）。作者建議以尺度因子校正並反向重走路徑。
作者的適用性主張
原文未報告
任務需求與來源
原文未報告來源：原文未報告
作者報告的結果
stationary range RMSE 0.001 m at the ~2 m target and 0.021 m at the ~15 m target over eight 5-min collections, no range walk despite vent temperature rising from 22.3 to 45.0 deg C (Sec. 4.1)
error did not grow between consecutive targets or with target offset (Sec. 4.2, Fig. 3)
限制
原文未報告
inter-target distance RMSE 0.033-0.128 m across nine circuits, largest (0.128 m) for the single two-floor circuit (Table 3)
RMSE correlates linearly with cumulative distance travelled (R2 0.63-0.80); scale factor and reverse-path repetition suggested (Sec. 4.2, Sec. 5)
failures to initialise, mostly with visual SLAM disabled; mid-scan orientation and alignment shifts only in multi-loop multi-floor circuits (Sec. 4.3)
repeating a path can push error beyond the SLAM correction bounds and cause catastrophic misalignment that is hard to fix without timestamps (Sec. 5)
平台與感測器（原文）
handheld；LiDAR in Leica BLK2GO (830 nm; FOV 360 deg h x 270 deg v; range 0.5 to 25 m; 420,000 pts/s)；3-camera panoramic vision system in BLK2GO (4.8 Mpixel, 300 x 135 deg, global shutter) used for visual SLAM
coarse registration median TE 0.053 m and RE 0.272 deg over the successful cases of 250 perturbed samples (Table 1)
fine registration median TE 0.0246 m vs 0.0422 m for point-to-point ICP (Table 2)
100% coarse success for alpha_r >= 0.5 deg and alpha_t >= 0.4 m, unlike all ten baselines (Sec. 4.1.2, Fig. 10)
on seven floors of an active site, fine-registration errors of 0.009-0.203 deg and 0.032-0.107 m relative to the CloudCompare reference, about 0.05 deg and 0.058 m on average (Table 4)
per-point association separates rebar, glass, barriers, temporary materials, pipes and boxes from built structures and flags unbuilt decoration walls (Sec. 4.2.2, Fig. 15)
限制
registration accuracy only; association assessed qualitatively; GCP count small (3-4 per floor)
drift errors in the FAST-LIO2 reconstruction caused some wall points to receive low association levels (floors 07 and 12); authors suggest a high-precision laser scanner (Sec. 4.2.2)
geometry-only association: barriers close to walls can be associated with the wall even with normal verification (Sec. 4.2.2)
only regular human-made structures are modelled; MEP objects are not included (Sec. 4.2.2)
offline only; point clouds with few planar segments cannot be registered (Sec. 4.2.2)
coarse registration slower than PLADE and RANSAC in simulation (64.2 s vs 8.85 s and 27.27 s) (Sec. 4.1.2)
coarse precision drops for sparse or partial clouds from early construction stages (Table 3)
(inference) the real-site reference is itself a manual CloudCompare registration seeded by 3-4 GCPs, so centimetre-level differences are not independently verified
(inference) assumes gravity axis known and planar primitives dominant
平台與感測器（原文）
handheld；real site: handheld sensor suite with Ouster OS0-128 LiDAR (clouds built with FAST-LIO2)；simulation: ISPRS indoor modelling benchmark clouds from stationary, handheld and backpack scanners
幾何profile-wise systematic shifts X/Y/Z and RMSE (Results tables)任務原文未報告
場域與營建關聯長距離坑道（地下基礎設施類）；證明僅兩端控制時長距離 SLAM 隧道測量會產生系統誤差，對隧道施工測量控制布設具直接意義（推論）。
作者主張、報告結果與限制
研究重點
作者在 Josef 礦坑主坑道 750 m 直線段測試 Emesent Hovermap ST-X，以 Leica P40 與 MS60 建立毫米級測量網與球靶，比較單程（1P）與往返（2P）各五次掃描，逐剖面計算橫向、縱向與垂直方向的系統偏移與 RMSE。結果顯示短段精度高，但全長出現累積誤差與縱向壓縮，需尺度改正，橫向偏差可達數十公分；往返掃描可降低橫向誤差但系統誤差仍存在。
作者的適用性主張
adequate for many applications; geodetic precision over long distances needs scaling and path optimization
任務需求與來源
原文未報告來源：原文未報告
作者報告的結果
local cloud accuracy ~7 mm average RMSEL, no local deformations (Conclusion; Tab. 2)
double pass roughly halves the transverse deviation: average RMSDOT 206.4 mm (1P) vs 115.4 mm (2P) with S=1 (Tab. 2; Fig. 5)
similarity (scaled) transformation reduces longitudinal error (ØRMSEΔY 24.0 to 10.3 mm for 1P, 30.9 to 4.8 mm for 2P) but leaves transverse and vertical shifts almost unchanged (Tab. 2); the Conclusion states deviations are roughly halved
限制
single device; mining gallery
cumulative error over 750 m with systematic longitudinal compression (clouds about 10 cm per km too short; scale 1.00007-1.00018) requiring scale correction (Results; Tab. 1; Conclusion)
transverse error up to 430 mm single pass and still > 100 mm with double pass; vertical shifts not reduced by double pass (Conclusion)
vertical shift about 90 mm on average (ØRMSEΔZ 89.3-92.8 mm) regardless of pass type; authors suggest possible sagging from smooth floor versus rough rock ceiling, to be verified (Tab. 2; Conclusion)
sign of lateral error depends on walking direction (Conclusion)
平台與感測器（原文）
backpack；Emesent Hovermap ST-X (32-channel LiDAR, range 0.5-300 m, FOV 290 x 360 deg, up to 640,000 pts/s single return)；reference: Leica ScanStation P40 TLS + Leica Nova MS60 robotic total station
機器人資料蒐集與掃描規劃3D digitization via multi-robot exploration
場域類型
受控實驗原文未報告
平台與感測器
輪式地面機器人四足機器人商用 SLAM 掃描儀光達（研究系統）RGB-D 或深度相機
參考量測
作者所稱的 ground-truth 地圖是遙控機器人執行 SLAM 的產物（未說明是哪一台），不是獨立參考；文中未報告點雲幾何精度（期刊版 Sec. 4.1、6）。原文未報告
量測指標
幾何原文未報告任務原文未報告
場域與營建關聯案例為大學校園內約 80 m²、以雜物與障礙物分隔為兩區的空間，含半完成（清水混凝土與 CMU 牆）與已完工的實驗室及辦公室，非施工中工地，作者於結論也稱其為雜亂的實驗室環境；所稱 ground-truth 地圖是遙控機器人執行 SLAM 的產物（文中未說明是哪一台機器人），並非獨立參考（VoR Sec. 4.1、6）。
作者主張、報告結果與限制
研究重點
作者提出多代理人（機器人與人員）營建資料蒐集方法：以 IFC 網格水平切片得到 2D 地圖，將已探索自由面積占地圖自由面積的比例作為邊界探索的停止準則，未達門檻而已無可達邊界時向其他代理人求助。案例在大學校園約 80 m² 的空間進行，含半完成（清水混凝土與 CMU 牆）與已完工（實驗室及辦公室）區域；因無 BIM，改以遙控機器人執行 SLAM 預先建立的 2D 地圖作為面積基準；RA1（Robotnik SUMMIT-XL，搭載 Emesent Hovermap ST 蒐集點雲、Ouster OS1 供導航）探索到 84.53% 時已無邊界，RA2（Boston Dynamics Spot 加機械手臂）在人員遠端指定抓取點後移除障礙，RA1 再探索至 95.79% 完成。文中未報告點雲幾何精度、密度或執行時間。
作者的適用性主張
原文未報告
任務需求與來源
原文未報告來源：原文未報告
作者報告的結果
exploration stopped with no frontiers at 84.53%; after RA2 removed the obstacle it reached 95.79% and met the criterion (Sec. 4.3)
only one human intervention needed (indicating the grasp point) (Sec. 4.3)
MANET radio up to 120 Mbps with configurable frequency, power, bandwidth, 3 x 3 MIMO and AES-256 encryption (Sec. 4.2.1, Table 2)
限制
原文未報告
small case study: two robots, one human, one task; a bigger and more complex scenario is needed (Sec. 5)
ROS 1 master node and centralized CMS are single points of failure (Sec. 3.1.1, 5)
prior information (BIM or pre-mapped ground truth) needed for robust and complete results (Sec. 5)
only one agent performed exploration (Sec. 5)
communication stability depends on space layout and wall materials (Sec. 5)
human needed to indicate the grasp point (Sec. 4.3, 5)
tested in a cluttered lab, not a real construction site (Sec. 6)
平台與感測器（原文）
wheeled UGV (Robotnik SUMMIT-XL, RA1)；legged (Boston Dynamics Spot, RA2)；human agent；Emesent Hovermap ST LiDAR scanner (data collection; range up to 100 m; frames registered by its LiDAR-based odometry)；Ouster OS1 3D LiDAR (navigation; 150 m maximum range, 45 deg VFOV)；front-facing short-range depth camera (RA1 perception)；RGB-zoom camera and thermal camera on a 3-axis gimbal, five short-range depth cameras (RA2)
以 BIM 或參考地圖定位與對齊進度與跨期變更long-term map alignment to BIM or TLS, map extension and change detection (construction monitoring)
場域類型
施工中工地公開基準1 construction site (ConSLAM), 4 sequences
平台與感測器
人員攜帶（手持、背包、穿戴）光達（研究系統）IMU
參考量測
評估參考是最終 ICP 所用的同一張 ConSLAM TLS 地圖；與 ConSLAM 提供的真值軌跡僅作差異比較。not independent: the evaluation reference is the same TLS map used for the final ICP
量測指標
幾何APE RMSE of trajectories versus TLS-referenced poses; registration RMSE and fitness at 1 cm and 3 cm任務qualitative positive and negative difference maps (Figs. 13-14)
construction site monitoring and other domains needing fast-updated 3D maps (abstract)
任務需求與來源
原文未報告 (authors state ground-truth poses ideally need about 1 cm accuracy, Sec. 6)來源：原文未報告
作者報告的結果
Automatic alignment without manual initialization of the first keyframe, unlike ground-truth generation in ConSLAM or Newer College (Sec. 7)
Alignment to a clutter-free BIM despite scan-map deviations from clutter, furniture and dynamic objects (Sec. 1, 6)
Not restricted to Manhattan-world layouts; supports extending the reference map (Sec. 7)
限制
as-built BIM derived from TLS; reflections; initial-pose sensitivity; offline runtime
Sensitive to the initial SLAM or LIO poses; large drift, especially Z-drift in narrow corridors without floor or ceiling points, may not be corrected (Sec. 8, Fig. 15)
Large deviations of permanent walls or columns, low overlap or symmetric environments can defeat alignment; the final ICP may be wrong where the reference map is wrong (Sec. 8)
Not real-time; final ICP can take several dozen minutes (Sec. 8)
Indoor Scan Context needs a 360-degree horizontal FoV, so solid-state LiDARs and depth cameras are not directly supported (Sec. 8)
Window reflections create fictitious elements in change detection (Sec. 8)
Only clutter and dynamic-object deviations are addressed, not alterations of permanent building elements (Sec. 1)
After the KNN loops, rotational APE rises in S3 and S5 (Table 1, Sec. 6); Sec. 7 explains this pattern (naming it for sequences 2 and 5 there) by erroneous KNN loops detected where the ISC-aligned trajectory, before Umeyama alignment, still deviated about 1.5 m in Z and X from the ground truth; the final ICP filters these loops (Sec. 6, Sec. 7, Table 1)
Correct ISC correspondences are very sensitive to the number of top candidates (N_c = 100) (Sec. 5.2.2)
(inference) Final-ICP poses to the TLS map serve as ground truth, so the TLS-referenced final stage is not independently evaluated
(inference) The ConSLAM BIM was modelled from the TLS point cloud of sequence 2 (Sec. 5.1), i.e. an as-built model, so design-versus-as-built discrepancy (brief Sec. 11 risk 2) is not tested; the same caveat is recorded for stuhrenberg2025liobim in C11b
平台與感測器（原文）
handheld (ConSLAM sequences)；3D LiDAR of the ConSLAM handheld system (model not reported; the ISC descriptor requires a 360-degree horizontal FoV)；9-axis IMU of the ConSLAM handheld system (model not reported; used for DLIO deskewing; LiDAR-IMU extrinsics from OA-LICalib)
authors claim efficient collection of high-quality point clouds for building interiors
任務需求與來源
minimum point spacing thresholds s_v,lim = s_h,lim = 0.013 (units as in Table 3) from cited data quality requirements來源：cited point data quality requirements [53] (not verified)
completeness wall 72.05% and floor 47.26% vs 61.05% and 30.86% for the comparative scenario (Table 5)
minimum densities column 0.001501, floor 0.01664, wall 0.005396 per mm2 vs 0.000829, 0.00143, 0.002773 for the comparative scenario (Table 5)
with the enhanced DWA SLAM mapping ran smoothly, while the conventional DWA made mapping fail between positions 3 and 4 (Sec. 4.1)
限制
completed buildings only; accuracy vs reference not verified
improper robot motion can cause point cloud mismatching due to limited sensor FoV (Sec. 1, Fig. 1)
only visual odometry used for LiDAR pose estimation; sensitive to lighting, blur and shadows (Sec. 5)
no independent geometric reference (TLS or survey) is used in the full text
column minimum density 0.0015 /mm2 below the GSA 0.00188 /mm2 requirement in poorly lit bottom regions (Sec. 4.2)
floor completeness below 50% because clutter blocked robot access; table tops not scanned due to sensor height (Sec. 4.2)
chair IoU only 62.55%, attributed to training and test chair geometry mismatch; segmentation needs fully labelled custom data (Sec. 4.3, 5)
(inference) the 'ablation' comparator changed platform (UGV), SLAM package (RTAB-Map) and motion integration at once, so gains cannot be attributed to motion-SLAM integration alone
平台與感測器（原文）
legged (Unitree Go1)；wheeled UGV in the comparative scenario (model not reported)；solid-state LiDAR RGB-D camera Intel RealSense L515 (HFOV 70 deg, VFOV 50 deg, angular resolution 0.07 deg)；2D LiDAR Slamtec Mapper (navigation)
GeoSLAM heavy noise near narrow passage; Livox FoV gaps; actuated Velodyne uniformly higher noise (Sec. 3)
M3C2 distance SD to TLS above 200 mm for all systems (364, 281, 232 mm), attributed to drift without global positioning; outliers deliberately kept (Sec. 3, Table 2)
Livox double wall and double floor in cross-section A-A' and ROI 2 (Sec. 3)
drift compensation needs reference data and serves evaluation only, not mapping (Sec. 2.5)
critical metric thresholds are application-dependent (Sec. 4)
(inference) single site and a single pass per system
平台與感測器（原文）
handheld；GeoSLAM ZEB Horizon (300,000 pts/s, 100 m, 30 mm at 100 m)；in-house Velodyne VLP-16 on a Dynamixel servomotor, 3D printed handheld rig, no IMU；in-house Livox Horizon (240,000 pts/s, 90 m, 20 mm at 25 m) with internal IMU, hand-carried；reference TLS RIEGL VZ-400i (60 stations, plane-patch registration SD 2.6 mm)
authors limit claims to static environments and call for BIM updating on sites
任務需求與來源
原文未報告來源：原文未報告
作者報告的結果
overall translation RMSE 0.080 m and rotation 0.663 deg; 34% translation improvement over baseline ICP (Table 4)
semantic filtering alone (Sem (ORG)) improved overall translation error by 18% over ICP (ORG) (Sec. 4.3)
Z drift (Delta Z) of the BIM-based localization stayed within 0.084 m on the four tested sequences, against up to -0.968 m (LOAM), -2.675 m (DLO) and -1.547 m (Open3D SLAM) (Table 6)
assumes static built environment; BIM updates needed for dynamic construction sites (Sec. 4.5)
as-designed vs as-built deviations can cause failure (Sec. 4.5)
depends on previous pose; no global localization; floor and initial pose set manually at the first frame (Sec. 4.5)
reference trajectory from Cartographer SLAM run in 2D (x, y, yaw) with tuned parameters, not an external survey (Sec. 4.3)
authors state that the BIM-based localization does not show better 2D pose accuracy than the LiDAR-only LOAM, DLO and Open3D SLAM on the four tested sequences; Table 6 is mixed: lowest translation RMSE on Seq. 3-3 but highest yaw RMSE on Seq. 2-3, 4-2 and 5-2 (Table 6, Sec. 4.4)
axis-aligned Dynamo bounding boxes cause mixed or wrongly labelled map points (Fig. 8, Sec. 4.5)
long, narrow corridors with only walls and floors degrade tracking and caused a trajectory discontinuity on the 4th storey (Sec. 4.5)
平台與感測器（原文）
原文未報告: Fig. 5 shows a Velodyne VLP-16 with portable power and a laptop; the carrier is not stated; data collected with planar motion (Sec. 4.1, 4.4)；3D LiDAR (Velodyne VLP-16)
作者在 iOS 手持裝置上以 Apple ARKit 內建的 SLAM 將 BIM 疊合於施工現場，並在一棟地下一層至地上六樓、樓地板面積 280 m² 的 RC 施工中建築實測，發現掃描路徑中斷、移動過快造成影像模糊、施工中表面不平，以及現場材料與設備變動，都會使模型疊合偏移。作者因此提出操作程序：沿地面連續掃描的封閉路徑並在角落設校正參考物、平均移動速率約 0.25 至 0.5 m/s 並以特徵點數提示、以整平的動態參考平面作為 BIM 放置基準，並每 1 至 3 個工作天重掃特徵點地圖。單次移動速率試驗中，1.00 m/s 時虛實物件中心相距 7.9 cm，0.5 m/s 以下為 4.3 至 5.2 cm。系統另以顏色在 AR 中顯示構件進度落後、如期或超前。
作者的適用性主張
improved BIM overlay positioning on construction sites (abstract)
任務需求與來源
原文未報告來源：原文未報告
作者報告的結果
virtual-to-physical centre offset 4.3 to 5.2 cm at 0.50 m/s or slower vs 7.9 cm at 1.00 m/s (Table 2)
no positioning hardware needs to be mounted on site; relies on ARKit feature maps (Sec. 2)
five AR progress scenarios demonstrated on site (Sec. 7)
限制
未查證
feature-point mapping affected by poor lighting and occlusion (Sec. 8)
effectiveness on large projects and integration with project management software not tested (Sec. 8)
feature-point map becomes invalid after about 1 to 3 working days of site change and must be rescanned (Sec. 4.2.4)
(inference) accuracy evidence limited to one movement-rate test with a single measurement per rate; no before-and-after comparison of the proposed procedure
平台與感測器（原文）
handheld (iOS mobile device; model not reported)；camera and motion sensors of an iOS handheld device accessed through Apple ARKit (visual-inertial; device model not reported)
Consistent estimates through sensor dropout and recovery on two excavators (abstract, Sec. VI)
Translation consistency error mean 0.58 cm, SD 3.16 cm during a navigation task, defined as deviation between propagated and optimized estimates (Sec. V, Fig. 6)
Relative position error against RTK GNSS in the Construction Task: mean 0.18 cm, SD 1.28 cm, the smallest SD of the three estimators (Table I)
Latency 41 us mean versus 102 us for TSIF and 32.9 us for MSF (Table II)
Map built from the estimates stayed intact near trees where the TSIF-based map was corrupted by GNSS loss (Fig. 5, qualitative)
限制
Precise extrinsic calibrations are difficult to obtain on large machines and can degrade fusion (Sec. V)
Joint estimation of extrinsics and sensor time offsets left to future work (Sec. VI)
平台與感測器（原文）
walking excavator (two Menzi Muck machines; HEAP is one of them)；IMU at 100 Hz (Ellipse-A at the bottom of the cabin on the first machine; Lord MicroStrain MV5 on the roof of the second)；3D LiDAR Ouster OS0-128 (odometry from CompSLAM at 5 Hz)；dual-antenna RTK GNSS Leica iCON iXE3 at 20 Hz；discrete cabin-rotation joint angle measurements
地下與坑道測繪SLAM 點雲精度評估underground mine survey of extracted surfaces
場域類型
地下或隧道摘要未提及
平台與感測器
人員攜帶（手持、背包、穿戴）掃描器型式未說明或未查證
參考量測
摘要稱以 TLS 資料驗證（僅讀摘要）。TLS
量測指標
幾何摘要未提及任務摘要未提及
場域與營建關聯愛沙尼亞地下油頁岩礦開採後表面的測量情境（關鍵字 Oil shale mine、Estonia）；致謝提到國營企業 Eesti Energia 的主管與專家提供在 Estonia 地下礦改進測量方法的機會，兩台掃描儀由測量公司 Hades Geodeesia OÜ 提供；與地下工程出渣或超挖量測概念相近（推論）；非施工工地。
作者主張、報告結果與限制
研究重點
作者評估手持 SLAM 掃描用於地下礦場測量與開採後表面 3D 建模，並以 TLS 資料驗證；典型差異在水平與垂直方向分別約 2 cm 與 5 cm 以內。作者也指出傳統礦場測量結果受測量人員主觀影響，認為 SLAM 手持掃描最適合地下礦場測量。
typical discrepancies within 2 cm (horizontal) and 5 cm (vertical) vs TLS (Abstract)
限制
摘要未提及
平台與感測器（原文）
handheld；two SLAM-enabled handheld laser scanners provided by Hades Geodeesia OÜ (models not stated on the abstract page)；TLS reference；conventional mine surveying (compared; instruments not stated on the abstract page)
幾何range standard deviation for coloured targets (Table 1); no geometric accuracy of tunnel clouds任務leakage areas summarized for the 50 m DB section (Table 2); qualitative visual agreement with RGB images
authors state the workflow extracts leakages quickly and efficiently
任務需求與來源
原文未報告來源：原文未報告
作者報告的結果
saturated concrete targets showed 70 to 80% lower intensity than dry; dry targets above 27, saturated at or below 11 (Sec. 4.4, 7)
range SD against best-fit plane stayed between 9.2 and 13.0 mm for all non-black targets up to 6 m (Table 1)
ambient lighting changed intensity by 3 units or less (Sec. 4.3)
11 leakage areas with location and area extracted in the 50 m DB section; extracted surfaces showed good visual agreement with RGB images (Table 2, Sec. 5.2, 7)
限制
no quantitative leakage ground truth; intensity depends on range, incidence and colour
two small false leakage detections on patch repairs (Sec. 5.2, Fig. 18c, Sec. 7)
validation only visual against RGB images; no quantitative leakage ground truth or hydraulic head information (Sec. 5.2)
black targets gave larger range noise, up to +/-26.9 mm SD (Table 1)
only one walking speed tested and two 50 m sections; longer sections and multi-epoch linking left to future work (Sec. 5.1, 7)
(inference) intensity depends on range, incidence and surface colour, requiring calibration per device
平台與感測器（原文）
handheld (field scans walked at 0.5 m/s along the track centre)；static placement for laboratory target scans (30 s each)；commercial SLAM scanner Emesent Hovermap with rotating Velodyne VLP-16 puck (360 deg FoV, 905 nm); IMU and SLAM algorithm not described in the article
以 BIM 或參考地圖定位與對齊AR inspection / remote robot control localization in BIM
場域類型
已完工建築原文未報告
平台與感測器
人員攜帶（手持、背包、穿戴）光達（研究系統）IMU
參考量測
唯一參考是每段錄製起點以人工量測的初始位置；誤差同時包含配準誤差與 SLAM 漂移，漂移的貢獻未量化（Sec. 4.1、4.3）。評估只用手持錄製，Spot 機器人僅作示範。原文未報告
量測指標
幾何localization error (0.03 m; 0.2-0.3 m)任務原文未報告
場域與營建關聯應用動機為建築進度控制、數位輔助維護與遠端巡檢；測試在 TU Wien 研究室旁約 28 m 長、2.5 m 寬的走廊，以及 TU Wien 圖書館六樓的環形走廊進行，兩者都是既有建築；圖書館樓層可見 BIM 未記載的家具、植物與牆體、門窗差異，且無法進入個別房間；非施工中工地（Sec. 1、4.2、5）。
env. 1 (28 m x 2.5 m hallway, hallway-only BIM): XY error 0.03 m and Z 0.035 m (medians averaged over keyframes); registration successful at all keyframes of 10 recordings (Sec. 4.3)
env. 2: registration virtually always successful from keyframe 28; mean XY 0.19 m and Z 0.24 m up to keyframe 36 (Sec. 4.3)
supports non-perpendicular walls and varying floor levels, unlike Herbers and Konig (Sec. 5)
registration runs in the background without degrading live tracking (Sec. 4.3)
限制
原文未報告
self-similar floor plan causes wrong template matches for up to 20 first keyframes; success only 30% in keyframes 2-10 and 50-90% in keyframes 12-26 (Sec. 4.3, 5)
error grows with accumulated SLAM drift, notably after keyframe 38 (about 60.8 m); afterwards XY below 0.3 m and Z below 0.4 m (Sec. 4.3)
ground truth only for the initial sensor position, measured manually; SLAM drift contribution not quantified (Sec. 4.1)
differences between BIM and building (furniture, plants, missing or extra walls, doors, windows) hamper registration (Sec. 5)
only corridors on one floor recorded; rooms not accessible (Sec. 4.2)
平台與感測器（原文）
handheld: custom wooden frame with dual hold and 24 V battery (all evaluation recordings)；legged robot Boston Dynamics Spot (demonstration setup, not evaluated)；Ouster OS0-128 Gen 2 LiDAR (128 channels, 90 deg VFOV; 512x20 in env. 1, 1024x20 in env. 2; up to 131,072 points per frame) with IMU data from the sensor
變形、地工與土方監測SLAM 點雲精度評估geotechnical monitoring: convergence and rockfall detection
場域類型
地下或隧道含任務層驗證2 mines
平台與感測器
車輛或工程機具人員攜帶（手持、背包、穿戴）商用 SLAM 掃描儀
參考量測
FARO static TLS; survey targets; field confirmation of convergence
量測指標
幾何absolute/relative trueness and precision (median, MAD), drift vs trajectory length, SLAM intrinsic/extrinsic precision任務detection of simulated rockfall (down to 2.5 x 5 cm in structured scene) and mine convergence
作者在一座營運中塊狀崩落法礦場與科羅拉多礦業學院 Edgar 實驗礦，以 Kaarta Stencil 2 與 Emesent Hovermap 兩款 SLAM 行動掃描比對 FARO 靜態掃描，建立非參數（中位數、MAD）的絕對與相對精度、SLAM 內在、外在精度與密度覆蓋指標。結果顯示迴圈閉合可大幅降低漂移，以 SLAM 將新期資料配準到基準圖可取得遠優於絕對精度的相對精度，並實證以 3 cm 偵測門檻偵測岩塊掉落與最多約 10 cm 的收斂變形。
作者的適用性主張
SLAM MLS provides data quality required to detect geotechnically relevant changes
任務需求與來源
LoD 0.05 m wall-to-wall convergence; 0.1 m rockfall來源：derived by authors from cited geotechnical literature (Sec. 2.1)
作者報告的結果
target-level absolute trueness within the sensor's +-3 cm: maximum mean 1.58 cm and median 1.65 cm (Stencil 2); Hovermap mean 0.75 cm and median 0.72 cm (Sec. 4.4, Tables 7-8)
relative target-level accuracy with M3C2: mean magnitude 0.22 to 0.58 mm, MAD 2.9 to 6.7 mm (Sec. 4.5, Table 9)
loop closure reduced drift from up to 1.0 m horizontal and 0.5 m vertical to at most 0.25 m and under 5 cm over the first 240 m (Sec. 4.6, Fig. 10)
loop closure lowered intrinsic SLAM sigma from 7.11 cm to 2.86 cm (Table 6)
SLAM-based registration: median 0.4 cm, MAD 2.9 cm (Stencil 2 with Mine Vision Systems software, to static) and median 0.05 cm, MAD 1.36 cm (Hovermap to Hovermap, Emesent software) (Sec. 4.6)
MLS 30 s vs static 18 min for a 45 m section; about five times higher effective sampling rate (Sec. 3.3, Table 3)
MLS wall density about 8000 to 10,000 pts/m2 at 10 km/h with more uniform coverage than static scans (Sec. 4.3)
限制
mining context; vendor black-box SLAM; later sections not fully read
error distributions non-Gaussian with heavy outliers on rough surfaces (Sec. 4.1)
entry-aligned drift reached 0.18 m at 150 m (0.12% of trajectory); global ICP hides this drift; at 0.12% drift, segments longer than about 25 m exceed the 3 cm sensor bound (Sec. 4.4, Sec. 5.2)
opposite-direction passes gave larger discrepancies (MAD 10.30 cm) than same-direction passes (4.26 cm) (Sec. 4.5, Table 6)
C2M produced false positives from mesh gaps and normal orientation errors on rough surfaces; the smallest rock was not identified with C2C in the unstructured scene (Sec. 4.7, Figs. 13-14)
site-level overfitting is a concern where drifts shift globally in the rock mass (Sec. 5.2)
dynamic mine operations make repeatable survey trajectories difficult (Sec. 5.1)
more work needed on optimal collection and processing (Abstract)
場域與營建關聯於真實建築工地以靜止機器人測試三個位置（每處約 1 分鐘、約 300 次光達掃描，結果為三次執行的平均），現場有移動工人、木板與設備箱等雜物，並以全測站追蹤機器人上的稜鏡作為參考，且依實測參考牆偏差修正參考值。表 II 至 IV 的模型偏差是作者在網格中把上下兩側結構人為拉開 0.3 m 所模擬，並非實測施工偏差；表 I 才是未加人為偏差的雜物影響比較。屬少數在施工中環境以全測站為獨立參考、在建築平面圖模型中定位的研究（Sec. IV）。
作者主張、報告結果與限制
研究重點
作者主張施工中牆體缺漏、臨時物與實作偏差使 ICP 對整棟 BIM 的對位不可靠，因此提出「局部參考」：先對整個平面圖模型做點到平面 ICP，再只對選定的參考牆面（至少三個互不平行的面）精修，並以影像密度估計網路的分數剔除或加權雜物、人員等離群點後融合到光達點。實驗在真實建築工地以靜止機器人搭配移動工人與雜物進行，並以全測站追蹤機器人上的稜鏡作為參考，且修正參考牆的竣工偏差；表 II 至 IV 的模型偏差是作者把網格上下兩側結構人為拉開 0.3 m 所模擬。
作者的適用性主張
outlier filtering and measurement selection are key for on-site precise localization
任務需求與來源
原文未報告 (motivated by on-site robotic fabrication accuracy)來源：原文未報告
作者報告的結果
Lowest position RMSE per location came from selective localization with semantic information: 232 mm (filtered) vs 390 mm for full-model ICP at location A, 68 mm vs 222 mm at B, 52 mm (weighted) vs 76 mm at C, i.e. at least 30% lower error (Tables II-IV; Abstract)
Selective localization against reference walls constrains two directions well (trace close to the maximum eigenvalue), while full-model ICP is uncertain in more than one direction (Sec. V, Fig. 6)
Only on-board sensing, no markers or site preparation (Sec. I)
限制
static robot; plan-derived mesh; small sample
no single method combination always worked; semantic filtering performance location-dependent (Sec. V)
lateral uncertainty high where few walls constrain one axis (Sec. V)
robot stationary during evaluation; ~1 min / 300 scans per location (Sec. IV-C)
Tables II-IV use a model whose upper and lower structures were moved 0.3 m apart to simulate a severe deviation; only Table I (157 to 218 mm at A and 74 to 83 mm at B with clutter) uses the unmodified plan (Sec. IV-C)
High failure rates for some variants: 76.5% (selective, weighted) and 50.3% (selective, full) at A, 52.6% (selective, weighted) at B, 25.6% (selective, filtered) at C (Tables II-IV)
Binary semantic filtering at location C removed nearly all points on the lateral reference wall; the density network, trained on NYU indoor data, partly filtered building structure outside its training domain (Sec. V)
Plan-derived mesh with a planar floor and equal wall heights because no floor or ceiling information was available (Sec. IV-B)
feasibility of real-time UAV-UGV interaction demonstrated (Highlights)
Blimp followed the manually driven UGV through a 115 s hallway trial covering about 21 m at 0.2 m/s (Sec. 5.2, Fig. 21)
All modules ran in real time on one laptop during a 3 min test (Sec. 5.1)
限制
full text not reviewed
tested in construction-like indoor environment rather than an active site (Abstract)
Validation targets integration and real-time operation, not the accuracy of RTAB-Map, Vins-Mono or LNSNet; drift was judged only from test videos (Sec. 5, 6)
Relative positions from marker detection and SLAM comparison differ by roughly 25 cm (x, y) and 30 cm (z), attributed to marker error (about 6 cm) and sensor offsets (Sec. 5.2)
Blimp is blown off course near air-conditioning vents and fans; no UAV obstacle avoidance (Sec. 5, 6)
UAV cannot follow the UGV above 1 m/s because movement commands arrive at 0.5 Hz (Sec. 5.2)
2D LiDAR and stereo maps are not integrated (Sec. 4, 6)
平台與感測器（原文）
wheeled UGV built on a Clearpath Husky A200 (Sec. 3.1)；custom indoor helium blimp, 1.83 m ellipsoid envelope, 300 g payload (Sec. 3.2)；stereo camera ZED (forward-looking, on a pan-tilt unit) feeding LNSNet segmentation and RTAB-Map (Sec. 3.1, 4.1, 4.2)；2D LiDAR HOKUYO URG 04LX-UG01 at the UGV front for a 2D occupancy map; scan limited to 180 deg of 240 deg, 4000 mm maximum radius (Sec. 3.1, 4)；wide-angle camera Point Grey Flea3 on a tilt unit tracking a Whycon marker on the blimp (Sec. 3.1, 4.3.1)；UGV wheel encoders fused with an IMU by EKF, used as odometry for RTAB-Map (Sec. 4.2)；blimp: Raspberry Pi camera v2 (rolling shutter, 410 x 308 at 30 fps) and Bosch BNO055 9-DOF IMU (100 Hz) for Vins-Mono (Sec. 3.2.1, 4.3.2)
91% of deviations below 5 cm after outlier removal in the tower closed loop (ground floor to top floor and back) and in the museum basement closed loop (Sec. 5.1, 5.2.1)
about 1 m3 pillar sample: 97% (ZEB-REVO RT) and 100% (LiBackPack C50) below 5 cm (Sec. 5.3.1-5.3.2)
ZEB-REVO RT noise 3.8 mm vs LiBackPack C50 5.7 mm vs TLS 3.0 mm (Table 2)
mobile systems capture areas occluded from static stations more easily (Sec. 5.3.4)
限制
ICP alignment hides global drift; completed buildings
registration quality lower than static systems; operating mode crucial; large loops over several floors remain a limit (Sec. 7)
alignment by manual point picking + ICP to TLS, so absolute georeferencing error not isolated (Sec. 4)
two-floor loop through a narrow stairwell: 55% of deviations below 5 cm and 80% below 10 cm after outlier removal, 'accuracy' 9 cm (Sec. 5.2.2)
indoor-outdoor loop: ZEB-REVO RT captured only a limited part of the facade; 81% below 5 cm (Sec. 5.2.3)
low MMS point density prevented a proper comparison of the large laboratory floor; dataset 3 compared only on a ~1 m3 pillar (Sec. 5.3)
several percentages computed after outlier removal whose procedure is not described (Sec. 5.1-5.2)
平台與感測器（原文）
handheld；backpack；GeoSLAM ZEB-REVO RT handheld: 2D infrared laser profilometer with automatic rotating head plus IMU; no GNSS, no camera (Sec. 3.1, Table 1)；GreenValley LiBackPack C50 backpack: laser profilometer in rotating housing (360 deg H, 30 deg V FOV), panoramic camera, IMU, GNSS for outdoor positioning (Sec. 3.1, Table 1)
safe indoor navigation with dynamic obstacle avoidance demonstrated in the lab (Experimental Setup and Results, Conclusions)
Hector SLAM scaled to mapping large indoor construction spaces and its occupancy map supported autonomous navigation (Conclusions)
higher point density associated with better accuracy (Fig. 8)
限制
原文未報告
low resolution and accuracy of the low-cost 2D LiDAR (Experimental Setup and Results)
few, manually placed waypoints gave poor coverage; mid-room waypoints lowered density on far walls
SLAM drift caused incorrect registration of some local points
manual registration of the cloud to BIM lowered measured accuracy
initial occupancy map must be created by manual driving; site data were collected after working hours
平台與感測器（原文）
wheeled UGV (Clearpath Jackal)；two orthogonal 2D LiDARs (horizontal, 16 m range, for Hector SLAM mapping and navigation; vertical, 10 m range, for cross-section scans; models not reported)；wheel odometer and IMU of the Clearpath Jackal
Enables long-term, wide-area people tracking with a carried 3D lidar, field-tested with caregivers in a hospital (33 sequences, about 52 min) (Sec. Field test)
With the ground-plane constraint a 45 min, about 2400 m indoor sequence gave a flat, consistent map, whereas BLAM and LeGO-LOAM aborted before the end (Sec. SLAM framework evaluation, Fig. 3)
Offline mapping took 5382 s vs 15,327 s for BLAM on the same sequence (Table 1)
With GPS constraints a 42 min outdoor sequence with large undulations was mapped; without GPS no loop was found and mapping failed (Fig. 8)
Angular-velocity pose prediction kept localization working when the observer ran at about 3.0 m/s (Table 2, Fig. 13)
People positions agreed with a nine-camera Kinect v2 OpenPTrack setup with mean differences of 0.0768 m (observer) and 0.0990 m (subject) (Table 4)
限制
Ground-plane constraint assumes a single flat floor; moving between floors required manually switching maps (Sec. Ground plane constraint; Sec. Field test)
Without the plane constraint the indoor map was warped by accumulated rotation error; outdoors without GPS no loop was found and mapping failed (Sec. SLAM framework evaluation)
No trajectory ground truth: localization was evaluated as the difference between predicted poses and NDT results (Sec. Sensor localization evaluation)
Initial localization pose is given by hand; acceleration-based prediction made results worse because of noise (Sec. Sensor localization)
README parameter-tuning guide advises selecting the registration method (FAST_GICP recommended; FAST_VGICP or NDT_OMP when speed matters) and tuning ndt_resolution (0.5-2.0 m indoor, 2.0-10.0 m outdoor) to obtain good odometry
Independent construction-site evaluation reports limited front-end odometry accuracy compensated by back-end graph optimization with loop and ground constraints (feng2025_construction_lidar_eval, Sec. 5.3)
平台與感測器（原文）
backpack with 3D LiDAR and PC carried by a walking or running human observer (indoor building floors, hospital, outdoor sloped terrain)；3D LiDAR (Velodyne HDL-32e, 360 deg range data at 10 Hz); its angular velocity output is used for UKF pose prediction during online localization；GPS (outdoor mapping only, unary position edges in UTM coordinates; model not reported)
先以無人機（UAV）影像經運動恢復結構（SfM）產生粗略現況點雲，轉為可通行網格與體素地圖，以射線追蹤與貪婪覆蓋選出地面機器人的最佳靜態掃描站位與路徑。機器人以 2D Hector SLAM 提供粗配準，再以 ICP 精配準停走式掃描。作者以商用地面雷射掃描（TLS）點雲為參考，並比較整體作業時間。
作者的適用性主張
authors state both TLS and robot achieved satisfactory accuracy and the robot was more time-efficient
任務需求與來源
scan completeness (LoC) and registration without targets; no tolerance stated來源：原文未報告
作者報告的結果
RMSE of GRoMI cloud vs commercial TLS 3.91 cm (Sec. 5)
total time for six scans 50 min vs 108 min for TLS workflow including 0 min manual registration (Table 4)
scan planning reached LoC < 4% in all 8 simulations (Table 3)
限制
single site, flat ground, TLS reference registration details not reported
UAV use restricted by regulations, weather and GPS interference (Sec. 6)
2D SLAM limits robot to flat or nearly flat ground; 3D SLAM planned (Sec. 6)
commercial TLS reported as slightly more accurate than GRoMI (Sec. 5)
moving-robot point cloud is noisier and used only for navigation; final map relies on stationary scans (Sec. 4.3)
平台與感測器（原文）
wheeled UGV；UAV；2D line laser scanners: 4 vertically mounted (60 m working range, 50 Hz, 190 deg vertical line) for 3D mapping and 1 horizontal for Hector SLAM；DSLR camera (RGB mapping)；UAV camera with 1/2.3-inch CMOS sensor (SfM prior map, 30 m altitude, 80% overlap, GSD 1 cm)；navigation view camera and object-avoidance sensors (navigation only)
["correct transformation ranked first on the three simulated datasets and second on both real datasets (Tables 3 to 5, Sec. 6)", "correct transformation close to the reference before fine registration: eps_R at most 0.12 deg and eps_T at most 181.1 mm (Table 7)", "centroid support reduces congruent bases by a further factor of 200 or more (Sec. 5.3)", "4-PlCS ranks the correct transformation at least as well as 3-PlCS and better on Steel-1 (Table 8, Sec. 5.7)", "handles symmetry, self-similarity and clutter via plane-patch support (Sec. 6)"]
限制
["4-plane bases give no clear computational benefit over 3-plane bases
4-PlCS was slower on four of five datasets (abstract, Table 8, Sec. 5.7)", "on real data several plausible candidates with similar support (e.g., a one-floor offset on UW-E5) require a final visual choice by the user (Sec. 5.4, Tables 4 and 5)", "Point Support reduces candidates by only 5 to 10% and does not improve the ranking while adding cost (Sec. 5.5, Table 6)", "coarse only
fine registration still required (Sec. 6)", "relies on sufficient planar structure (inference)", "ground-truth transformations for the real datasets are the result of fine registration, not independent survey (Sec. 5.2)"]
平台與感測器（原文）
["simulation", "real construction-site scans (acquisition platform not reported)"]；["real datasets UW-E5 and Mercury-1: laser scans from construction sites, scanner model not reported in the paper (Sec. 5.1, Acknowledgments)", "simulated datasets House-1, House-2, Steel-1: points generated on BIM mesh surfaces with sigma = 2 mm noise, no subsampling (Sec. 5.1)"]
機器人資料蒐集與掃描規劃多站掃描配準as-built 3D data collection and multi-scan registration (robotic reality capture)
場域類型
已完工建築受控實驗2 test beds (one building floor; one outdoor area)
平台與感測器
輪式地面機器人光達（研究系統）IMU相機
參考量測
室內以相鄰掃描中編號較高者為參考計算 RMSE；室外距離比對所用的量測方式未說明；最終點雲沒有 TLS 或全測站參考（Tables 3、5）。partial: consecutive-scan NN RMSE with the higher scan ID taken as reference (Table 3); outdoor SLAM inter-scan distances compared with measured 'actual distances' whose measurement method is not stated (Table 5); no TLS/total-station reference for the final cloud
量測指標
幾何inter-scan NN RMSE and axis deviation angles; SLAM vs measured distances (Tables 3, 5)任務registration success vs feature-based baseline (Table 4)
authors claim improved automation and fewer scans for large sites
任務需求與來源
原文未報告 (no tolerance stated)來源：原文未報告
作者報告的結果
SLAM-based registration succeeded where feature-based registration failed for low-overlap scan pairs (Table 3 vs Table 4)
fewer scans and no targets required for registration (Sec. 6)
限制
no active construction site; relative (not absolute) accuracy
dynamic (moving) SLAM point cloud noisier, lower resolution and RGB fusion not meaningful (Sec. 6)
tested indoors on a building floor and outdoors near buildings; real construction sites left to future work (Sec. 6)
(inference) 2D Hector SLAM assumes near-planar motion; unsuitable for multi-level or uneven terrain without 3D SLAM
平台與感測器（原文）
wheeled UGV；2D LiDAR (5 SICK 2D laser scanners: 1 horizontal used by Hector SLAM, 4 vertically mounted on a rotating frame; each -95 to 95 deg at 0.3333 deg increments)；IMU；wheel encoders (four wheels)；DSLR camera (RGB mapping of static scans)；navigation camera and object-avoidance sensors (navigation only)
at 45% to 70% overlap and 24 m to 43 m scan spacing ICP RMSE rose to 2.265 to 4.293 m while plane matching stayed at 0.092 to 0.132 m (Table 2)
Testbed #1 registered with three of six scans (#1, #4, #6), reducing the number of scans needed (Table 6)
final RMSE 0.198 m, 0.183 m and 0.047 m for Testbeds #1, #2 and #3 (Table 6)
no artificial targets or manual alignment (Abstract, Discussion)
限制
requires three planes with one intersection point in the overlapped area (Discussion and Conclusion)
only the single set of three largest planes and one corner point is used for fine registration (Final Alignment section)
ICP outperforms plane matching when overlap exceeds about 89% and scan distance is below about 10 m (Fig. 5, Table 2)
accuracy is measured against the higher-ID scan taken as ground truth, not an independent survey (Results)
(reviewer observation) Table 4 (0.427 deg) and Table 6 (0.637 deg) list final deviation angles above the < 0.35 deg and < 0.34 deg values in the abstract and text
平台與感測器（原文）
customized robotic hybrid LiDAR system whose base frame is a mobile robot platform; 360 deg scans (200 s each) at 6, 3 and 2 scan positions in the three testbeds；robotic hybrid 3D LiDAR system with four SICK 2D line laser scanners (80 m working range, 25 Hz scan speed, 200 s per 360 deg scan, 190 deg vertical line; 0.1667 deg vertical and 0.072 deg horizontal resolution; SICK model not stated)；regular digital camera on the scanner, eight images per 360 deg scan, used for RGB texture mapping through the mounting kinematics and a pinhole model
tower full raw cloud vs CRP: mean 0.025 m, SD 0.034 m; 67% of points within 2 cm (Table 5)
courtyard 12 x 7 m wall vs TLS DSM: mean 0.017 m, SD 0.023 m, 99% within 5 cm (Table 12)
ice house outward vs return after optimisation: 97% within 5 cm (Table 7)
rapid acquisition, e.g. 19 million points in 10 min in the tower (Sec. Test dataset presentation)
suitable for the 1:100 to 1:200 architectural scale (Conclusion)
限制
原文未報告
mine cave loop-closure error of almost 40 cm; raw outward vs return mean 0.214 m, SD 0.313 m (Table 8, Fig. 13)
2016 circular 660 m loop in open, irregular terrain drifted: mean 0.531 m vs UAV DSM, 0.578 m in the church area (Tables 16-17)
ZEB1 deviations increase with height because of the spring-driven ray distribution (Sec. courtyard (C))
edges and architectural details become rounded (Fig. 16)
no radiometric data; not adequate at 1:50 or larger scales (Conclusion)
results depend on manual segmentation, cleaning and filtering (Sec. Metric validation)
平台與感測器（原文）
handheld；GeoSLAM ZEB1 (spring-mounted head, 40 Hz, about 43,200 points/s, 15-30 m range, about 1.5 kg) (Table 1)；GeoSLAM ZEB-REVO (automatically rotating head, 100 Hz, 270 deg HFOV and 100 deg VFOV, about 2 kg) (Table 1)；both: Hokuyo UTM-30LX 2D 905 nm ranging sensor and an IMU with triaxial gyros, accelerometers and magnetometers; no GNSS and no RGB (Sec. ZEB system operational behaviour)
任務證據出處
Abstract
Rebolj et al. 2017 (point cloud quality for Scan-vs-BIM)
Quantitative, task-linked quality criteria per element size class (Sec. 3, Table 2)
Relation between criteria and scanning parameters for planning (Sec. 5)
Public research data (CC BY 4.0)
限制
Assumes clear line of sight; occlusions and construction interference treated as not relevant (Sec. 1.2)
Criteria depend on the identification algorithm (Sec. 4.4)
Future work needed on point cloud positioning and calibration, merging partial clouds and moving objects (Sec. 6)
Point cloud assumed pre-aligned to the BIM; registration and positioning errors not evaluated (Sec. 2.2)
Photogrammetry and videogrammetry validations reuse the simulated BIM that produced the criteria; only the Kinect 2 test uses a real scene with 18 elements (Sec. 4 intro, Sec. 4.3)
dmax is derived for a single frame; overlap from a moving scanner and merging of frames are not modelled (Sec. 5.1)
平台與感測器（原文）
simulation (virtual scanner moved along a modelled person trajectory in HeliOS)；Kinect 2 range camera scanning part of a real building (carrying mode not reported)；simulated person-carried mobile laser scanner with 360-degree vertical view (HeliOS)；synthetic images of the simulated BIM processed with VisualSFM (photogrammetry and videogrammetry)；Kinect 2 range camera
地下與坑道測繪現況與竣工建模maintenance model of difficult-to-access mine shaft
場域類型
地下或隧道原文未報告
平台與感測器
人員攜帶（手持、背包、穿戴）商用 SLAM 掃描儀
參考量測
沒有獨立的參考測量，品質只以下降與上升兩趟點雲的一致性評估（Sec. 4）。原文未報告
量測指標
幾何原文未報告任務原文未報告
場域與營建關聯礦場人員運輸電梯豎井（長約 1440 m、直徑 5 m），豎井須依國家標準定期檢查，該礦至少每三個月檢查一次；現行檢查需五人在電梯上、一人在地面以無線電聯繫，電梯停用約一小時；屬地下基礎設施維護與檢查情境，非施工中工地，也是早期 ISARC 以手持 SLAM 掃描難以到達區域的例證（Sec. 1、3）。
作者主張、報告結果與限制
研究重點
作者於芬蘭 Pyhäsalmi 礦場長約 1440 m、直徑 5 m 的 Timo 豎井，在例行檢查時站在以 1 m/s 移動的電梯車廂頂，以 ZEB1 手持 SLAM 掃描器（彈簧式 2D 雷射加 IMU）掃描，下降與上升各約 25 分鐘、各覆蓋約 270°，共取得 1.2 億點。GeoSLAM 雲端處理前兩次完全失敗，第三次才產出結果，且 252 至 434 m 段上下兩個半圓環未能正確疊合；作者以時間戳分離上下行點雲，分割五條纜線自動對齊後套用轉換，半手動修正該段。重疊部分距離差在自動 SLAM 段（130 m 深處）平均 27 mm、標準差 17 mm，修正段平均 37 mm、標準差 19 mm。全文沒有三腳架 TLS、全測站或標靶作為獨立參考；作者認為系統尚未成熟到能勝任此任務，但速度快，可作為維護模型的潛在資料來源。
作者的適用性主張
原文未報告
任務需求與來源
原文未報告來源：原文未報告
作者報告的結果
about 120 million points; total measurement time about one hour including preparation (Sec. 3-4)
estimated point density 2,778 points per m2 (about 2 cm spacing) at 1 m/s (Sec. 4)
automatic SLAM section at 130 m depth: down and up overlap distance error mean 27 mm, SD 17 mm (Sec. 4, Figs. 11-12)
results seemed not strongly affected by rain and condensation (Sec. 5)
限制
原文未報告
GeoSLAM SLAM processing failed completely twice before a third attempt gave a result (Sec. 4)
252-434 m section misregistered (two half-cylindrical rings did not overlap); automatic and manual point registration failed; fixed by cable-based alignment with mean 37 mm, SD 19 mm (Sec. 4, Figs. 5-10)
moving platform allowed only about 270 deg per pass; full 360 deg coverage recommended (Sec. 1, 5)
constant rain in the first 1000 m wetted the scanner window and rock walls (Sec. 3)
no independent reference survey; quality assessed only by down and up consistency (Sec. 4)
authors state the system was not yet ready for this demanding task (Sec. 5)
平台與感測器（原文）
handheld, operator standing on the roof of a mine elevator car moving at 1 m/s and waving the scanner up and down at about 1 Hz; scanner rotated 180 deg between the down and up passes；ZEB1 handheld SLAM laser scanner (commercial name; the whole system, developed by CSIRO, is called Zebedee): spring-mounted 2D laser, data logger and IMU; 43,200 points/s; typical range 15-20 m; noise +-30 mm; 270 deg horizontal FOV, about 120 deg swept vertical FOV
尺寸與平整度檢核concrete floor surface flatness (regularity) compliance control
場域類型
已完工建築含任務層驗證2 floor slabs in one building
平台與感測器
固定式（三腳架）地面光達（TLS）
參考量測
manual 2 m straightedge and precision steel rule on a chalk-line grid matching the generated straightedges (Sec. 7.2)
量測指標
幾何per-straightedge deviations and FF values versus manual measurement; sensitivity to point subsampling任務individual straightedge deviations and overall floor compliance compared with manual control (Fig. 12)
場域與營建關聯已完工建物中兩片約 25 年的混凝土實驗室樓板（Heriot-Watt 大學，6.40 m × 6.70 m 與 4.80 m × 8.10 m 區段），以人工直尺為獨立參考並比對合格判定；非施工中工地，也非 SLAM 點雲。是本批次紀錄中唯一具標準化判定與獨立參考的任務方法錨點；是否為全語料中唯一未經系統檢索確認（推論）。
TLS provides sufficient accuracy for standard surface regularity control (Sec. 9)
任務需求與來源
Straightedge maximum deviation tolerances and FF/FL F-Number tolerances來源：BS EN 13670:2009, CONSTRUCT NSCS, BS 8204, ACI 117-06, ASTM E1155-96 (Sec. 2.1, Tables 1-2)
作者報告的結果
Straightedge deviations from TLS agreed with manual measurements; mean differences 1.2 mm (SD 1.0) and 0.7 mm (SD 0.4) for the two slabs with no significant difference in a two-tailed t-test at alpha 0.05 (Fig. 12)
Using 4%, 10% or 25% of the initial point clouds did not materially change straightedge or F-Number results (Sec. 8.1.1, 8.3)
Many more straightedges than a manual survey could afford; equivalent manual Random or Grid-Star surveys estimated at about 35 h and 23 h (Sec. 8.2)
Results link to BIM objects and are repeatable by other stakeholders (Sec. 9)
With 230 straightedges each on the Acoustic Lab slab (320 each on the Drainage Lab slab), the Random and Grid-Star methods found maximum deviations of 11.3 and 11.4 mm versus 7.6 mm with Grid-Square on the Acoustic Lab slab, identifying a non-compliant area (100% global flatness tolerance 10 mm) that Grid-Square missed (Sec. 8.2, Fig. 13)
The whole TLS workflow took about 1 h 50 min and 1 h versus 3 h (17 straightedges) and 1.5 h (10 straightedges) for manual Grid-Square control (Sec. 8.1.2, Table 5)
限制
two slabs; TLS only; F-Numbers not validated against manual F-Number survey
Two straightedges showed notable manual-versus-TLS differences whose cause (manual or scanning error) is unclear (Sec. 8.1.1)
F-Numbers agreement with 3 m straightedge equivalences was weaker for one slab and needs further validation (Sec. 8.3, Sec. 9)
Only two aged laboratory floor slabs tested (Sec. 7.1)
Aligning TLS scans with the BIM may need user input (Sec. 3)
Scanner error figures are indicative and depend on material and incidence angle (Sec. 2.2)
(inference) Evidence is for static TLS; whether SLAM point clouds with centimetre-level noise and drift support the same decisions is untested
Half of the TLS workflow time was scanning, with high-accuracy settings about five times slower than standard settings (Sec. 8.1.2)
Grid-Square straightedges missed a non-compliant area that Random and Grid-Star detected, so a standard sparse layout can under-report defects (Sec. 8.2)
現況與竣工建模SLAM 點雲精度評估as-built BIM geometry creation (Scan-to-BIM) of a building corridor
場域類型
已完工建築含任務層驗證1 (single corridor)
平台與感測器
推車人員攜帶（手持、背包、穿戴）商用 SLAM 掃描儀相機
參考量測
Faro Focus3D TLS (12 scans, 32 tie points) georeferenced with a Leica TS15 total-station network adjusted in LGO; total-station element measurements
量測指標
幾何ICP RMS and cloud-to-cloud deviation (height function, least-squares planes) after ICP alignment to TLS任務door and window width and height differences between Revit models from each data source
場域與營建關聯已完工且使用中的建物（UCL South Cloisters 一樓走廊）之 BIM 幾何建立適用性測試，非施工中工地；屬早期（2013）以 SLAM 行動掃描做任務層比較（門窗尺寸）的非 MDPI 研究，是否為最早之一未經系統檢索確認（推論）。
suitable for asset capture and facility management; not for millimetre-level survey engineering or monitoring (Sec. 5)
任務需求與來源
原文未報告 (authors mention millimetre-level accuracy for survey engineering and monitoring as a benchmark, Sec. 5)來源：原文未報告
作者報告的結果
Residuals after ICP registration to TLS were within a few centimetres for both systems (Tables 1-2, Sec. 4.1.2)
Large time savings relative to 12 TLS scans taking about 5 hours including control survey (Sec. 3.2, Sec. 5)
Handheld form factor can reach areas such as stairwells that trolleys cannot (Sec. 2)
限制
single site; comparison after ICP alignment; manual artefact removal
Not adequate for applications requiring millimetre-level accuracy such as survey engineering and monitoring (Sec. 5)
i-MMS SLAM is 2D only, restricting use to areas without significant height change (Sec. 2.1)
Artefacts from people and glass were removed manually before comparison (Sec. 4.1.1)
Visible noise in the mobile clouds at detail level (Sec. 4.1.3, Fig. 7)
Wall thickness on the office side could not be measured by laser scanning because offices were inaccessible (Sec. 3.1)
Single corridor, single test (Sec. 3.1)
(inference) Point-cloud comparison was done after a best-fit rigid ICP alignment to the TLS cloud, so reported deviations reflect residual distortion and noise, not absolute georeferencing accuracy; drift is captured only insofar as it distorts the cloud relative to the rigid fit
平台與感測器（原文）
trolley；handheld；Viametris i-MMS trolley: three Hokuyo line scanners (270 deg swath; two produce the point cloud, one upright scanner feeds the 2D SLAM) and a Point Grey Ladybug spherical camera；ZEB1 handheld: the same Hokuyo scanner on a spring with a small IMU, oscillated by the operator, online 6-DoF SLAM；reference: Faro Focus3D phase-shift TLS (12 scans, 32 tie points), Leica Viva TS15 total station network adjusted in LGO 8.1
["registration quality at least as good as commonly used AEC/FM software, simpler and faster (abstract, Sec. 6)", "after ICP fine registration, results were similar or better than the point-based packages in 75 to 92% of the 12 registrations (Table 2, Sec. 5.2)", "fine registration converged in fewer iterations from the proposed coarse alignments (Sec. 5.2)", "using 10% of the points did not affect accuracy and kept computation low (Sec. 5.2)", "semi-automated one-click extraction stayed accurate on the Bicocca data where automated extraction failed (Sec. 5.3)"]
限制
["plane matching is manual (abstract, Sec. 4.3)", "fully automated model-scan registration in AEC/FM is complex and often ill-posed because of self-similarities, clutter and multi-object models (abstract, Sec. 3)", "fails when three non-parallel planes are not visible
one of the 12 scans failed because no horizontal plane was visible (Sec. 5.2, 6)", "automated scan plane extraction was poor and took about 3 h on distant Bicocca scans
w_plane is critical and smaller values raise run time about tenfold (Sec. 4.2.1, 5.3)", "cylindrical surfaces are not supported
octree acceleration and automated matching left for future work (Sec. 6)", "timing comparison is indicative only, with two users (Sec. 5.2)"]
平台與感測器（原文）
["terrestrial laser scanner (model not reported)；Engineering V scans of a concrete structure under construction and Bicocca scans of about 5 M points each (Sec. 5.2, 5.3)"]
作者指出直尺與剖面儀等傳統平整度檢查速度慢、量測稀疏且需接觸表面，因此提出以地面雷射掃描點雲偵測混凝土平整度缺陷。三種演算法（距離影像濾波 RF、偏差濾波 DF、滑動視窗 SW）都先以整體最小平方法擬合參考平面，差別只在降噪與偏差計算步驟。作者以含黏土模擬缺陷（直徑 3 至 50 cm、厚 1 至 7 mm）的水平試驗平板，定義偵測、定位與誤報三項指標，比較一台 AMCW 與兩台 TOF 掃描儀在 3 至 20 m 距離與不同角解析度下的表現。SW 偵測表現最好但有時定位較差；距離增加使表現下降，AMCW 下降較快。摘要所稱 20 m 可偵測 3 cm 寬、1 mm 厚缺陷，對應 Scanner 2 搭配 SW 的「偵測但未定位」結果（表 6）。
作者的適用性主張
laser scanners can be effectively used to assess surface flatness (abstract)
任務需求與來源
deviation from a flat reference beyond a specified tolerance (abstract, generic)來源：原文未報告
作者報告的結果
Objective evaluation framework (test bed, detection, localization and precision measures, test procedure) for comparing scanners and algorithms (Framework)
SW algorithm shows equal or better detection in all test cases (Conclusions; Fig. 4)
3 cm x 1 mm defect detected at 20 m by SW with Scanner 2 (TOF, 0.014 deg), though not localized (Table 6)
Performance maps support inspection-rate estimates: 0.346, 0.1 and 0.65 m2/s for Scanners 1-3 in a hypothetical slab task (Inspection Rates section)
限制
未查證
Only one-time scans of horizontal boards; no probability-of-detection statistics; non-horizontal surfaces untested (Discussion; Conclusions)
Within 6 m, noisy data at board boundaries merge defects with false alarms, especially for SW (Discussion; Fig. 6)
Beyond 6 m, shallow incidence causes occlusion and mixed pixels behind defects of 5 mm or thicker, elongating detections (Discussion)
Sparse, varying data density (4.3 cm spacing at 10 m) hampers the image-based RF and DF algorithms (Discussion)
SW sometimes has worse localization than RF and DF because it produces larger defect regions (Comparison of Algorithms)
Per a citing study, the defect metrics are not compatible with current standard flatness specifications (Straightedge, F-Numbers), making compliance assessment difficult (bosche2014flatness Sec. 2.2)
平台與感測器（原文）
static terrestrial (tripod)；three terrestrial panoramic laser scanners, models not named in the paper: Scanner 1 AMCW (FoV 360 x 310 deg; 0.036 and 0.018 deg tested), Scanner 2 TOF (about 5,000 points/s; 0.014 and 0.007 deg), Scanner 3 TOF (0.014 deg)
任務證據出處
abstract
Scan-vs-BIM object recognition and as-built dimensions
尺寸與平整度檢核dimensional compliance control of erected steel structure (column plumb, inter-column distances)
場域類型
施工中工地1 construction project, 5 scans
平台與感測器
固定式（三腳架）地面光達（TLS）
參考量測
none (authors state ground truth unavailable)
量測指標
幾何registration MSE and matched-point counts任務as-built minus as-designed column positions and plumb (no ground truth)
場域與營建關聯施工中鋼構廠房（Portland Energy Center 發電廠專案，多倫多）的現場 TLS 掃描，屬真實工地資料；以 AISC 303-05 與 MNL 135-00 為容許差來源示例（Sec. 3.2）。展示掃描誤差如何傳遞到竣工尺寸，但無獨立參考。
作者主張、報告結果與限制
研究重點
作者改良先前的方法，先以人工選三組以上對應點把工地雷射掃描粗對齊專案 3D CAD 模型，再以新的 ICP 精對齊整個模型，依與各構件表面相符的點數與覆蓋面積判定構件是否被辨識。接著對每個被辨識的構件個別再做 ICP，求得其竣工位姿，並與設計位姿比較，推算柱垂直度與柱間距等尺寸以檢查是否符合容許差。實驗使用加拿大多倫多一座發電廠鋼構廠房施工期間的五次掃描。作者坦承缺乏真值，且位姿偏差與掃描距離相關，結果尚不足以判斷尺寸合規檢查的精度。
作者的適用性主張
potential for automated as-built dimension calculation and control; accuracy not established (Sec. 3.3.1)
任務需求與來源
project dimensional tolerances, e.g. AISC 303-05 and MNL 135-00 (named as examples)來源：AISC 303-05, MNL 135-00 (Sec. 3.2)
作者報告的結果
Model fine registration lowered MSE and increased matched points for all five scans compared with the earlier method (Table 2)
Object fine registration further lowered MSE (13-37 mm2) (Table 4)
Recognition is quasi-automated, robust to clutter and occlusion, and efficient (Sec. 4)
The author states that recall improved for all scans; Table 3 shows higher recall for Scans 2-5 and equal recall for Scan 1 (83%), overall recall 83% vs 80% and precision 93% vs 91%, but lower precision for Scans 4 and 5 (93% vs 94%, 82% vs 84%); gains were small because the manual coarse registrations were already accurate (Sec. 2.4.3, Table 3)
Model fine registration took about 2 min per iteration on CPU for about 650,000 points and about 20,000 facets; without the acceleration it would take about two orders of magnitude longer (Sec. 2.4.2)
限制
no ground truth; range-dependent error; prefabricated-shape assumption
Assumes each object's shape already complies with tolerances, reasonable only for prefabricated elements (Sec. 3.1)
No ground truth for as-built poses or dimensions; results not reliable enough to conclude on accuracy for dimensional compliance (Sec. 3.3.1)
Calculated pose deviations correlate with scanner range (r = 0.45; columns 20-80 m from the scanner), possibly from range-dependent scanner error or fewer recognized points (Sec. 3.3.1)
Per-object registration is often ill-conditioned because column ends are occluded (Sec. 3.3.1)
Coarse registration is manual; mesh models lack semantics, so control points were computed manually (Sec. 2.1, 3.3.2)
Recall and precision rely on object presence identified by manual visual inspection of each scan (footnote to Sec. 2.4.3)
The expert manual time estimate (a few hours to one day) is not based on field measurements (footnote to Sec. 3.3.2)
Object poses are refined independently, which may produce clashes between objects (Sec. 5)
平台與感測器（原文）
static terrestrial (tripod)；terrestrial laser scanner (Trimble GX 3D per Sec. 1.1.2 and ref. [44])