Asadi et al. 2019 SLAM image-to-BIM registration
作者提出把影片關鍵影格即時對位到設計 BIM 的方法。定位部分以 ORB-SLAM2 為基礎，改用由影片產生的自訂詞袋字典提升缺乏特徵室內場景的追蹤；第一次蒐集時以 MVE 稠密點雲和 BIM 的手動對應角點求相似轉換，建立真實尺度的全域地圖，之後各次蒐集只在該地圖中重新定位。每個關鍵影格先由 SLAM 位姿產生對應的 BIM 視圖，再以梯度下降最小化影像與 BIM 視圖之間消失點距離與消失線夾角，得到精修位姿。作者在走廊與室內施工工地兩段影片驗證，並報告在 Jetson TX1 上的每影格計算時間。
本頁內容
Real-time registration of video keyframes to an as-planned BIM: an ORB-SLAM2-based monocular SLAM with a custom vocabulary relocalizes in a BIM-scaled global map, and each keyframe pose is refined by aligning vanishing points and lines with the rendered BIM view; tested in a hallway and an indoor construction site on a Jetson TX1.
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 (webcam, 1920 x 1080 at 30 fps, fixed focal length; model not reported) |
|---|---|
| 原文測試平台 | UGV with a webcam from Asadi et al. 2018d (the AutCon 96:470-482 system); platform model and drive type not restated in this paper |
| 狀態估計 | Augmented monocular SLAM built on ORB-SLAM2 with a custom DBoW2 vocabulary generated from the video; first session builds a global map that is scaled by a manual similarity transform to BIM, later sessions relocalize in this map with the stored scale; keyframe poses then refined by gradient-descent alignment of vanishing points and vanishing lines between the keyframe and the rendered BIM view (max 500 iterations, thresholds 1 pixel and 1 deg) |
| 資料關聯 | ORB features for tracking and relocalization; Canny edges and Hedau et al. vanishing point voting for keyframe perspective; BIM vanishing points computed directly from model geometry |
| 時間表示 | discrete keyframes |
| 去畸變 | 不適用 (camera only) |
| 迴圈閉合 | ORB-SLAM2 place recognition and relocalization; the perspective step is presented as reducing drift before loop closure |
| 全域最佳化 | none beyond ORB-SLAM2; refined keyframe poses overwrite the SLAM poses at the end, described as similar to global bundle adjustment after loop closure |
| 地圖表示 | sparse ORB-SLAM2 map in real-world scale; dense MVE point cloud used once for manual alignment to BIM |
| 先驗資訊 | as-planned BIM (visible model lines, rendered views), manual first-frame registration via corresponding corners between dense point cloud and BIM |
| 可輸出幾何 | keyframe camera poses in the BIM coordinate system and the matching BIM views; no point-cloud accuracy product |
| 計算需求 | NVIDIA Jetson TX1 on the UGV: 657 ms (hallway) and 1,904 ms (construction site) average per keyframe; desktop Intel i7 3.4 GHz 6-core below 0.2 s per site keyframe |
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 |
|---|---|---|---|---|---|
| 相機 | webcam (monocular; model not reported) | 方法輸入 | 未標示 | 1,920 x 1,080 video at 30 fps; fixed focal length; intrinsics from MATLAB calibration; keyframes downsampled to 640 x 360 for real time | (Asadi et al., 2019, Sec. Experimental Setup and Results (Initial Setup)) |
| 載具平台 | UGV from Asadi et al. 2018d (model not restated) | 方法輸入 | 未標示 | moved along a path while recording video | (Asadi et al., 2019, Sec. Experimental Setup and Results (Initial Setup)) |
| 運算硬體 | NVIDIA Jetson TX1 | 執行運算平台 | 未標示 | runs the proposed SLAM and perspective processing on the UGV | (Asadi et al., 2019, Sec. Initial Frame Registration and Camera Localization; Sec. Computation Time) |
| 運算硬體 | desktop with Intel i7 processor | 執行運算平台 | 未標示 | 3.4 GHz, 6 cores | (Asadi et al., 2019, Sec. Discussion) |
作者報告的優勢與限制
優勢
- Augmented SLAM with custom vocabulary tracked two 90 deg turns in a featureless hallway where ORB-SLAM2 produced two separate hallways (Fig. 7; Sec. Initial Frame Registration)
- Angular errors were below 3 deg before and below 0.01 deg after iteration in both scenes (Figs. 9-10 text)
- Real-time registration for the hallway video (0.65 s per keyframe versus 0.74 keyframes per second) (Discussion)
限制
- On the cluttered construction site the 640 x 360 perspective estimation was inaccurate for about 15 keyframes; higher resolution was more accurate but far too slow (Discussion; Figs. 13-14)
- Construction-site processing (1.9 s per keyframe) exceeded the keyframe rate, so the UGV had to move slower (Discussion)
- Fine-pose alignment cannot correct the camera position along the direction normal to the image plane, which still relies on SLAM (Conclusion)
- Curved walls or arches are expected to produce higher error because the method relies on straight edges (Conclusion)
- First-frame registration and the first global map require a manual step (Method)
營建工程相關證據
其中一個驗證場景是施工公司提供 BIM 的室內施工工地（120 秒影片、59 個關鍵影格），用來說明影像對位 BIM 以支援進度監測；精度只以消失點與消失線的對齊誤差描述，沒有獨立量測的相機位姿參考（Experimental Setup and Results、Discussion）。
原文驗證環境：已完工建築、施工中工地
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 4 筆紀錄。
Asadi et al., 2019 · Text Computation Time 本方法 2 筆
指標average computation time per keyframe
表格設定（擷取紀錄原文）：Average computation per keyframe at 640 x 360 (rough pose from the augmented SLAM, VP/VL estimation, iterative fine-pose alignment) on the UGV's Jetson TX1; hallway video 70 s with 52 keyframes (Asadi et al., 2019, Text Computation Time)
average computation time per keyframe，authors' hallway video · hallway (52 keyframes)
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 Asadi 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 |
|---|---|---|
| proposed augmented SLAM + perspective alignment本方法原文提出硬體：NVIDIA Jetson TX1 on the UGV | 657 ms | (Asadi et al., 2019, Sec. Computation Time (Fig. 12 values stated in text)) |
Asadi et al., 2019 · Text Discussion 本方法 1 筆
指標processing time per keyframe on desktop (stated as less than 0.2 s)
資料集與序列authors' construction-site video · indoor construction site
表格設定（擷取紀錄原文）：Same construction-site keyframes processed on a desktop computer (Asadi et al., 2019, Text Discussion)
processing time per keyframe on desktop (stated as less than 0.2 s)，authors' construction-site video · indoor construction site
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 Asadi 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 |
|---|---|---|
| proposed method on desktop本方法原文提出硬體：desktop Intel i7 (3.4 GHz, 6 cores) | 0.2 s僅報告範圍註記（擷取紀錄）：other: upper bound stated in text ('less than 0.2 s') | (Asadi et al., 2019, Sec. Discussion) |
Asadi et al., 2019 · Text Practical Implications 本方法 1 筆
指標distance error of 18 pixels (Fig. 13) in the 17th keyframe
資料集與序列authors' construction-site video · indoor construction site, keyframe 17
表格設定（擷取紀錄原文）：Distance error of 18 pixels for keyframe 17 of the construction-site video, cited in the text with reference to Fig. 13 (per-keyframe mean square error between estimated and ground-truth vanishing points at 640 x 360) to explain the minor misalignment shown in Fig. 15(b and d) (Asadi et al., 2019, Text Practical Implications)
distance error of 18 pixels (Fig. 13) in the 17th keyframe，authors' construction-site video · indoor construction site, keyframe 17
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 Asadi 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 |
|---|---|---|
| proposed method (VP estimation at 640 x 360; keyframe 17 of the construction-site video)本方法原文提出 | 18 pixel | (Asadi et al., 2019, Sec. Practical Implications (refers to Fig. 13)) |
來源
Asadi et al., 2019
（2019）Real-Time Image Localization and Registration with BIM Using Perspective Alignment for Indoor Monitoring of ConstructionJournal of Computing in Civil Engineering, 33(5):04019031
DOI 10.1061/(asce)cp.1943-5487.0000847
同儕審查已出版已讀全文近十年查證後修正
相關版本
- conference precursor：Real-time image-to-BIM registration using perspective alignment for automated construction monitoring, Construction Research Congress 2018, pp. 388-397 (cited as Asadi and Han 2018; not read) https://doi.org/10.1061/9780784481264.038