BIM-Loc (S05)
BIM-Loc 以設計階段 BIM 作為先驗，將受差異影響的定位問題拆為 BIM 輔助軌跡最佳化與階層式差異偵測兩個耦合子問題，並迭代求解。其以多次命中射線投射建立點雲與 BIM 面的資料關聯，於位姿圖中加入掃描間一致性與掃描對 BIM 一致性因子，並以貝氏核推論在 BIM 表面紋理空間中逐像素、面、構件更新差異狀態。作者在模擬、已完工辦公建物（SLABIM）與施工中工地（CityU）評估，並明言其差異偵測只判斷構件存在與否，無法量化偏差大小。
本頁內容
Discrepancy-aware LiDAR localization that fuses odometry with BIM surface constraints in a pose graph and incrementally labels BIM elements as consistent, discrepant or unknown via kernelised Bayesian inference in texture space, validated in simulation, a completed office and an active construction site.
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 | 3D LiDAR (Velodyne VLP-16 in simulation; Livox Mid-360; Ouster OS0-128)、IMU、camera (visualisation only) |
|---|---|
| 原文測試平台 | simulation、handheld |
| 狀態估計 | Pose graph optimization with BIM-integrated factors solved incrementally with iSAM2 (GTSAM); front-end odometry is DLO (front-end agnostic) |
| 資料關聯 | Multi-hit ray casting against BIM facets; point-cluster plane factors (eigenvalue-based inter-scan BA-style) and point-to-BIM-surface residuals |
| 時間表示 | discrete poses |
| 去畸變 | 原文未報告 (delegated to front-end odometry) |
| 迴圈閉合 | none (drift bounded by BIM constraints) |
| 全域最佳化 | Online pose graph with odometry, inter-scan consistency and scan-BIM factors |
| 地圖表示 | As-designed BIM meshes (LOD 300, IFC) plus 2D texture-space discrepancy maps |
| 先驗資訊 | As-designed BIM (IFC, LOD 300, non-structural entities such as MEP and furniture filtered) plus an approximate initial pose (Sec. 6). In the simulation benchmark the initial pose was given in advance to all methods (Sec. 4.1.2); in the initial-pose sensitivity study on CityU Livox Mid-360 data the nominal initial poses came from the global scan-to-BIM registration of Zhang et al. (2024), which typically reaches 5-7 cm translation and within 1 deg rotation error (App. C.2) |
| 可輸出幾何 | BIM-aligned trajectory, aggregated scans, structure-level discrepancy labels (consistent, discrepant, unknown); no quantitative deviation magnitudes (Sec. 6) |
| 計算需求 | CPU only on a Mini-PC with Intel Core i9-12900 integrated with the sensor suite (Sec. 4); incremental iSAM2 in GTSAM (Sec. 5.1); batch-wise processing with 1.5 s batches (15 frames at 10 Hz): about 350 ms per batch for multi-hit ray casting, factor generation and discrepancy detection in parallel threads, and about 22 ms per frame for trajectory optimization (App. D); module means in Table 8 |
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 | Velodyne VLP-16 | 資料集感測器 | BIM-robot simulation benchmark | simulated sensor on a mobile robot in the Gazebo BIM-robot simulator | (Zhang et al., 2026, Sec. 4.1) |
| LiDAR | Livox Mid-360歸入：Livox MID-360 | 資料集感測器 | SLABIM (HKUST office benchmark) | handheld sensor suite with built-in IMU | (Zhang et al., 2026, Sec. 4.2, Fig. 13) |
| LiDAR | Livox Mid-360歸入：Livox MID-360 | 資料集感測器 | CityU construction benchmark | Floor 06 and Floor 08 sequences; each sequence has LiDAR scans with synchronized IMU measurements | (Zhang et al., 2026, Sec. 4.2, Sec. 4.2.1, Fig. 13) |
| LiDAR | Ouster OS0-128 | 資料集感測器 | CityU construction benchmark | handheld sensor suite; seven sequences, Floors 06 to 12 | (Zhang et al., 2026, Sec. 4.2.1, Fig. 13) |
| 慣性量測單元（IMU） | Livox Mid-360 built-in IMU | 資料集感測器 | SLABIM (HKUST office benchmark) | 原文未報告 | (Zhang et al., 2026, Sec. 4.2, Fig. 13 caption) |
| 慣性量測單元（IMU） | IMU (model not reported) | 資料集感測器 | CityU construction benchmark | synchronized IMU measurements included in each CityU sequence; the IMU device (built-in or external) is not named | (Zhang et al., 2026, Sec. 4 (inputs include IMU measurements), Sec. 4.2) |
| 相機 | camera (model not reported) | 資料集感測器 | CityU construction benchmark | part of the CityU sensor suite, used for visualization only (BIM overlay checks) | (Zhang et al., 2026, Sec. 4.2.1, Sec. 4.2.4, Fig. 17) |
| 運算硬體 | Mini-PC with Intel Core i9-12900 | 執行運算平台 | 未標示 | CPU only, no GPU acceleration; integrated with the sensor suite | (Zhang et al., 2026, Sec. 4) |
作者報告的優勢與限制
優勢
- ["Worst-case ATE on SLABIM office sequences 0.147 m translation and 2.911 deg rotation, lower than Fast-Loc and PALoc (Table 3)", "Lowest scan-to-BIM RMSE on all CityU construction sequences (Table 5)", "BIM storage far smaller than sampled point clouds (Sec. 5.2)", "In simulation, translation ATE RMSE stays at or below 0.046 m with up to 35% missing structures and up to 25% occlusion by extra structures shifted by up to 0.5 m, lower than DLO in all six tiers (rotation not lower in all tiers) (App. C.1, Table 9)", "Scan-to-BIM RMSE converges within 10 to 30 s over 50 perturbed initial poses (sigma 0.05 m and 3.33 deg) on CityU Livox Mid-360 data (App. C.2, Fig. 21)", "Texture-space discrepancy map needs about 5.0 MB for a 4000 m2 floor versus about 823.4 MB for a TSDF voxel map (Sec. 5.3)"]
限制
- ["Requires an approximate initial pose (Sec. 6)", "Discrepancy detection only for presence or absence of elements
- shape changes, boundary shifts and thickness variations not considered
- cannot quantify deviation amounts (Sec. 6)", "Requires BIM of at least LOD 300
- LOD 200 degrades accuracy
- LOD 100 unsuitable (Sec. 5.4)", "(inference) On the construction benchmark no ground-truth trajectories exist
- scan-to-BIM distance and WD use the same BIM as the prior, so they measure consistency with the design model rather than independent geometric accuracy", "Exploration is passive
- operators get no real-time coverage feedback, which risks incomplete or redundant scanning (Sec. 6)", "Degenerate when fewer than 3 to 4 non-coplanar structural facets are observed, e.g. facing a single flat wall (App. C.1)"]
營建工程相關證據
作者以 CityU 施工中樓層（06 至 12 樓）資料集評估（作者註明資料來自同團隊 Zhang et al., 2024），含手持 Ouster OS0-128 的 7 條序列與 Livox Mid-360 的 2 條序列，共 9 條、總長逾 3.5 km；工地含未完成結構、材料與臨時設備。該工地無真值軌跡，評估僅用 MME、scan-to-BIM 距離與以同一 BIM 為參考的 WD，以及影像疊合定性比對；以真值 ATE 呈現的優勢僅見於已完工辦公建物（SLABIM）。差異偵測的量化結果（F1）僅在模擬中取得。
原文驗證環境：模擬、公開基準、已完工建築、施工中工地、任務層驗證
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 5 個比較組，合計 49 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 1 組列在最後，並連到性能比較頁。
Zhang et al., 2026 · Table 3 本方法 18 筆
表格設定（擷取紀錄原文）：HKUST office (SLABIM F03-F05), handheld Livox Mid-360; GT trajectories from SLABIM; baselines use points sampled from the same BIM (Zhang et al., 2026, Table 3)
ATE RMSE translation，SLABIM (HKUST office benchmark) · 3F-Region1
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 Zhang et al., 2026 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(Zhang et al., 2026, Table 3)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| BIM-Loc本方法原文提出 | 0.147 m | (Zhang et al., 2026, Table 3) |
| Fast-Loc | 10.75 m | (Zhang et al., 2026, Table 3) |
| PALoc | 0.359 m | (Zhang et al., 2026, Table 3) |
Zhang et al., 2026 · Table 4 本方法 9 筆
指標RMSE of scan-to-BIM distance errors (truncated at 0.2 m)
表格設定（擷取紀錄原文）：HKUST office; per-scan point-to-BIM distance RMSE, distances < 0.2 m only; reference is the prior BIM itself (Zhang et al., 2026, Table 4)
RMSE of scan-to-BIM distance errors (truncated at 0.2 m)，SLABIM (HKUST office benchmark) · 3F-Region1
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 Zhang et al., 2026 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(Zhang et al., 2026, Table 4)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| BIM-Loc本方法原文提出 | 0.052 m | (Zhang et al., 2026, Table 4) |
| Fast-Loc | 0.067 m | (Zhang et al., 2026, Table 4) |
| PALoc | 0.077 m | (Zhang et al., 2026, Table 4) |
Zhang et al., 2026 · Table 5 本方法 9 筆
指標RMSE of scan-to-BIM distance errors (truncated at 0.2 m)
表格設定（擷取紀錄原文）：CityU active construction site, handheld LiDAR; no GT trajectories; scan-to-BIM RMSE (< 0.2 m) against the prior BIM (Zhang et al., 2026, Table 5)
RMSE of scan-to-BIM distance errors (truncated at 0.2 m)，CityU construction benchmark · Floor-06 (Mid-360)
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。
- 未執行
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 Zhang et al., 2026 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(Zhang et al., 2026, Table 5)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| BIM-Loc本方法原文提出 | 0.036 m | (Zhang et al., 2026, Table 5) |
| Fast-Loc | 0.08 m | (Zhang et al., 2026, Table 5) |
| PALoc | 0.087 m | (Zhang et al., 2026, Table 5) |
| LIO-BIM | 無數值未執行註記（擷取紀錄）：未執行 | (Zhang et al., 2026, Table 5) |
| CAD-Mesher | 無數值未執行註記（擷取紀錄）：未執行 | (Zhang et al., 2026, Table 5) |
Zhang et al., 2026 · Table 8 本方法 8 筆
資料集與序列CityU construction benchmark (ablation context, App. C.1)
表格設定（擷取紀錄原文）：Ablation (App. C.1): mean runtime of each BIM-Loc module with and without the discrepancy module; module given in metric_as_written; the unit of work (frame or batch) is not stated in the table; App. C.1 presents Table 8 together with Fig. 19 (CityU Floors 06 and 08, Livox Mid-360) (Zhang et al., 2026, Table 8)
Mean runtime duration of module (Multi-Hit Ray Casting)，CityU construction benchmark (ablation context, App. C.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 Zhang et al., 2026 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(Zhang et al., 2026, Table 8)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| BIM-Loc (w/ discrepancy detection)本方法原文提出硬體：Mini-PC with Intel Core i9-12900 CPU, CPU only (Sec. 4) | 114.92 ms | (Zhang et al., 2026, Table 8) |
| BIM-Loc (w/o discrepancy detection)本方法原文提出硬體：Mini-PC with Intel Core i9-12900 CPU, CPU only (Sec. 4) | 140.11 ms | (Zhang et al., 2026, Table 8) |
其他比較組
列出其餘 1 個比較組
來源
Zhang et al., 2026
（2026）BIM-Loc: BIM-integrated discrepancy-aware LiDAR-based indoor localizationThe International Journal of Robotics Research, OnlineFirst (article 02783649261462593; volume and pages not yet assigned)
線上優先已讀全文近十年查證後修正