BIM-Tracker
BIM-Tracker 以建築模型作為地圖，對影像序列做以模型為基礎的視覺追蹤，因此不需要迴圈閉合，誤差也不會累積。每一影格先依前一位姿以 Blender 光線追蹤找出 BIM 中可見的邊，把模型邊依長度取樣成三維點並反投影到影像，再沿垂直方向搜尋 Canny 邊緣建立 3D 對 2D 對應；每個取樣點保留兩側各一個假設，以 MSAC 剔除錯誤對應後用 Gauss-Newton 最小化重投影誤差估計相機位姿，最後以等速度卡爾曼濾波預測下一影格位姿。作者以 Zeb1 掃描建立走廊 BIM，用八組不同解析度、視野、遮擋與動態模糊的照片級合成序列量化誤差，並以智慧型手機真實影片示範擴增實境疊合。
本頁內容
Drift-free model-based visual tracking that registers each smartphone frame to edges rendered from a low level-of-detail BIM (MSAC with two hypotheses, Gauss-Newton pose refinement, constant-velocity Kalman prediction); quantified on eight photo-realistic synthetic corridor sequences and demonstrated on real smartphone video.
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 | monocular camera (smartphone Motorola G 1st generation for real data; virtual cameras for synthetic data) |
|---|---|
| 原文測試平台 | handheld smartphone (held near eye level, landscape)、simulation (photo-realistic rendered image sequences) |
| 狀態估計 | Model-based visual tracking: per frame, Gauss-Newton minimisation of reprojection errors between sampled 3D model edge points and image edges inside an MSAC framework (two hypotheses per sampled point, jump-out rules, iterative correspondence updates), followed by a constant-velocity Kalman filter that predicts the next pose; measurement covariance from error propagation of the least-squares solution (Sec. 3.7-3.9) |
| 資料關聯 | visible BIM edges rendered by Blender ray tracing (BVH), sampled into 3D points, back-projected and matched to Canny edges by searching perpendicular to the projected model edge (Sec. 3.6 states 25 to 40 px as empirically sufficient at 640 x 480 and 30 FPS, while Table 1 lists d_search = 100 px for the 640 x 480 synthetic sets); inter-frame correspondences reused when motion is small (Sec. 3.2-3.6, 3.10) |
| 時間表示 | discrete frames with constant-velocity Kalman prediction |
| 去畸變 | 不適用 (camera only) |
| 迴圈閉合 | none needed; each frame is registered to the model, so errors do not accumulate |
| 全域最佳化 | none |
| 地圖表示 | low level-of-detail 3D model derived from an IFC BIM (walls, floor, ceiling, doors), edges subdivided every 50 cm |
| 先驗資訊 | BIM of the corridor created manually from a Zebedee (Zeb1) point cloud; calibrated camera intrinsics; manually initialised first pose |
| 可輸出幾何 | 6-DoF camera trajectory in the BIM coordinate system; no point cloud produced |
| 計算需求 | about 0.89 s per frame in MATLAB on a desktop i7 at 2.6 GHz with 16 GB RAM (Table 4) |
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 |
|---|---|---|---|---|---|
| 行動掃描設備 | Zebedee (Zeb1) spring-mounted hand-held scanner | 方法輸入 | 未標示 | point cloud of the corridor used to model the BIM manually (prior map source); reported absolute accuracy 3 to 40 cm | (Acharya et al., 2019, Sec. 4.1) |
| 相機 | Motorola G 1st generation smartphone camera | 方法輸入 | 未標示 | 640 x 480, 30 fps, FOV 56.32 deg, sensor 3.63 x 2.72 mm; calibrated beforehand | (Acharya et al., 2019, Sec. 4.1.2; Table 2) |
| 運算硬體 | desktop computer, i7 processor at 2.6 GHz, 16 GB memory | 執行運算平台 | 未標示 | MATLAB implementation without hardware optimisation | (Acharya et al., 2019, Sec. 4.4) |
| 其他 | Agisoft PhotoScan Professional | 參考或真值量測 | 未標示 | bundle adjustment reference trajectory for real data with manually provided 3D coordinates; reprojection error 0.36 px, reconstruction error 4.65 mm | (Acharya et al., 2019, Sec. 4.3) |
| 其他 | virtual cameras rendered in Blender | 資料集感測器 | photo-realistic synthetic corridor dataset | sensor 32 mm x 24 mm; FOV 60, 90 or 120 deg; 320 x 240 to 1280 x 960; 30 FPS | (Acharya et al., 2019, Sec. 4.1.1; Table 1) |
作者報告的優勢與限制
優勢
- Average translational errors of 12.74 to 28.22 mm and rotational errors of 0.08 to 0.22 deg on synthetic sequences, 17.13 mm and 0.12 deg for the baseline set (Table 3)
- No accumulation of error along the trajectory, unlike ORB-SLAM on the same data (Figs. 11-13, 15)
- Robust to occlusions and motion blur in the synthetic experiments (Sec. 4.2.4)
- Real smartphone trajectory stays within the navigable corridor and supports AR banner projection (Sec. 4.3, Fig. 14-15)
限制
- Cannot recover after complete loss of tracking and needs a manual initial pose (Sec. 3.1, 5)
- Heavy motion blur, long turns, narrow FOV near walls (one or no visible model edge) and sharp rotations cause failures (Sec. 5)
- Inaccurate 3D models (wrong wall, floor or roof locations) can cause tracking failure (Sec. 5)
- About 0.89 s per frame in MATLAB, not yet real time (Sec. 4.4)
- Systematic errors where few building structures are near the camera view, and dense clutter such as stair railings (Sec. 4.2.5, 6)
營建工程相關證據
屬竣工後建築的室內定位研究：以 BIM 作為定位地圖，說明低細節度設計模型即可支援公分級相機定位；BIM 本身由 Zeb1 手持掃描點雲人工建立，真實資料只有定性評估，並未涉及施工中工地（Sec. 4.1、4.3）。
原文驗證環境：模擬、已完工建築
報告的性能數據
以下是原文作者報告的性能數值（author-reported results），不是本研究重新量測的結果。每張圖只並列同一個比較組（comparison group，同一張表、同一組實驗設定）內的方法；不同比較組之間的數值不可直接比較，也不構成排名。
本方法共出現在 3 個比較組，合計 27 筆紀錄。
Acharya et al., 2019 · Table 3 本方法 16 筆
表格設定（擷取紀錄原文）：Photo-realistic synthetic sequences rendered in Blender along an approximately 30 m corridor trajectory at 1.6 m/s and 30 FPS; errors against exact synthetic ground-truth poses; translation error = average Euclidean distance over the trajectory, rotation error = single Euler angle of the rotation difference (Acharya et al., 2019, Table 3)
Average translational error，photo-realistic synthetic corridor dataset · Set 1 (90 deg FOV, 320 x 240)
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Acharya et al., 2019 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| BIM-Tracker本方法原文提出 | 28.22 mm | (Acharya et al., 2019, Table 3) |
Acharya et al., 2019 · Table 4 本方法 9 筆
資料集與序列BIM-Tracker experiments · per frame
表格設定（擷取紀錄原文）：Average time per localisation step for one frame; MATLAB implementation without hardware optimisation (Acharya et al., 2019, Table 4)
average time: Visible edge detection，BIM-Tracker experiments · per frame
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Acharya et al., 2019 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| BIM-Tracker (Visible edge detection)本方法原文提出硬體：desktop computer, i7 processor at 2.6 GHz, 16 GB memory (MATLAB) | 0.11 s | (Acharya et al., 2019, Table 4; Sec. 4.4) |
Acharya et al., 2019 · Text Sec.4.5 本方法 2 筆
資料集與序列real smartphone corridor dataset · selected frames
表格設定（擷取紀錄原文）：Region of convergence on real smartphone data: random initial poses within +-1.5 m and +-20 deg of manually estimated true poses for a few frames (Acharya et al., 2019, Text Sec.4.5)
combined translational deviation handled per frame (approximately)，real smartphone corridor dataset · selected frames
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Acharya et al., 2019 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| BIM-Tracker本方法原文提出 | 0.7 m有附註註記（擷取紀錄）：other: approximate bound ('approximately +-0.7 m') | (Acharya et al., 2019, Sec. 4.5; Fig. 16) |
來源
Acharya et al., 2019
（2019）BIM-Tracker: A model-based visual tracking approach for indoor localisation using a 3D building modelISPRS Journal of Photogrammetry and Remote Sensing, 150:157-171
DOI 10.1016/j.isprsjprs.2019.02.014程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 程式碼釋出：debaditya-unimelb/BIM-Tracker (MATLAB implementation linked in the abstract) https://github.com/debaditya-unimelb/BIM-Tracker
程式碼：https://github.com/debaditya-unimelb/BIM-Tracker（授權：GPL-3.0 (LICENSE file on master branch checked 2026-09-25)）。有公開程式碼不等於已被重現，也不代表目前版本與論文版本相同。