RGB-D Mapping (Henry et al.)
RGB-D Mapping 以 PrimeSense（等同 Kinect）RGB-D 相機建立室內稠密三維地圖。相鄰影格的配準採用作者提出的 RGB-D ICP：先以 FAST 特徵與 Calonder 描述子加上 RANSAC 求得初始位姿，並以重投影誤差篩選內點，再把固定的特徵對應與稠密點到平面 ICP 誤差合併，以 Levenberg-Marquardt 聯合最佳化；兩階段版本只在 RANSAC 內點不足時才執行 ICP，因此在黑暗或缺乏特徵的區域仍能對齊。系統以關鍵影格、位置預篩選與詞彙樹偵測迴圈，再以 TORO 位姿圖或加入 ICP 點對的稀疏光束法平差（SBA）做全域最佳化，最後把點雲整合為面元（surfel）地圖。
本頁內容
RGB-D Mapping builds dense indoor 3D maps from a Kinect-style camera by combining RANSAC-verified sparse visual features with dense point-to-plane ICP in a joint frame-to-frame alignment (RGB-D ICP), detecting loop closures with keyframes and a vocabulary tree, optimising globally with TORO or sparse bundle adjustment augmented by ICP point pairs, and fusing the result into a surfel map.
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 | RGB-D camera (PrimeSense active-stereo camera equivalent to the Kinect sensor, 640 x 480 registered image and depth at 30 fps; Sec. 1) |
|---|---|
| 原文測試平台 | handheld (camera carried by a person and generally pointed in the direction of travel; placed on a tripod at each marker in the ground-truth runs) |
| 狀態估計 | frame-to-frame RGB-D ICP: RANSAC over sparse visual feature points (re-projection-error RE-RANSAC refined by two-frame sparse bundle adjustment, or Euclidean-error EE-RANSAC with Horn's method), followed by Levenberg-Marquardt joint minimisation of the fixed feature associations and dense point-to-plane ICP terms (alpha = 0.5 or beta = 1000); a Two-Stage variant runs the ICP stage only when RANSAC inliers are at most phi; with fewer than gamma = 10 inliers the previous relative motion is used as a constant-velocity initial guess (Secs. 3.1.1-3.1.4) |
| 資料關聯 | hybrid: FAST keypoints with Calonder descriptors (OpenCV) matched and verified by RANSAC with a 2.0 pixel (re-projection) or 0.03 m (Euclidean) inlier threshold; dense nearest-neighbour point associations via a k-d tree on a source cloud downsampled by a factor of 10, normals from PCA of small neighbourhoods; feature associations are not recomputed during ICP (Secs. 3.1.1-3.1.4) |
| 時間表示 | discrete RGB-D frames; each alignment uses only two consecutive frames (Sec. 5) |
| 去畸變 | 不適用 |
| 迴圈閉合 | keyframes created when the accumulated rotation or translation since the previous keyframe exceeds a threshold; candidates prefiltered by estimated global pose (keyframes within a few metres) and by a vocabulary tree on Calonder descriptors (ROS implementation), then accepted when RANSAC finds enough geometrically consistent feature matches (Sec. 3.2.1) |
| 全域最佳化 | either TORO pose-graph optimisation run to convergence after each detected loop closure, or global sparse bundle adjustment (SBA) over camera poses and feature points in which ICP-derived point pairs (distance and normal-angle filtered) connect frames lacking RANSAC inliers; SBA run for 5 iterations, projections with error above 2.0 pixels removed, then 5 more iterations (Secs. 3.2.2, 3.2.3, 4.3) |
| 地圖表示 | dense coloured point cloud of the aligned frames; surfel map (location, orientation, patch size, colour, confidence as a 2D histogram of viewing directions) created as a post-process after global optimisation (Sec. 3.3) |
| 先驗資訊 | none; single camera with known parameters (Sec. 3.2.3) |
| 可輸出幾何 | globally optimised camera trajectory, dense coloured point cloud map and a surfel map (for example 28 million points combined into 1.4 million surfels, Sec. 4.4); export format 原文未報告 |
| 計算需求 | hardware 原文未報告; ROS nodes as separate processes (Sec. 3.1.4); mean alignment time 0.20 to 0.21 s per frame for the RANSAC variants and 0.48 to 0.57 s for RGB-D ICP (Table 3); final global optimisation 1.08 s with TORO and 9.73 s with SBA on the Fig. 5 sequence, SBA amortised 0.12 s per frame; surfel map update about 3 s per frame (Secs. 4.2-4.4) |
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 |
|---|---|---|---|---|---|
| LiDAR | SICK laser scanner (model not reported) | 參考或真值量測 | 未標示 | used with a standard 2D SLAM approach to build the 2D reference map of the Intel Labs loop | (Henry et al., 2012, Sec. 4.1; Fig. 4) |
| RGB-D 相機 | PrimeSense RGB-D camera (equivalent to the sensor in the Microsoft Kinect) | 方法輸入 | 未標示 | 640 x 480 registered image and depth points at 30 frames per second; active stereo with an IR emitter and IR camera; RGB-D cameras described as having depth typically less than 5 m, noise about 3 cm at 3 m and a field of view of about 60 deg | (Henry et al., 2012, Sec. 1 (p. 647); Secs. 3.1.1, 4.5; Acknowledgements) |
| 其他 | 16 ground-truth markers with measured inter-marker distances and a tripod | 參考或真值量測 | 未標示 | consecutive marker distances 3 to 5.5 m; camera placed on a tripod at each marker | (Henry et al., 2012, Sec. 4.2) |
作者報告的優勢與限制
優勢
- Re-projection-error RANSAC averages nearly twice as many inliers per frame as Euclidean-error RANSAC (116.7 versus 60.3) and gives lower error on Intel-Day (Sec. 4.2; Tables 1-2)
- RGB-D ICP has lower mean marker-distance error than either RANSAC or ICP alone on both day and night sequences (Table 2)
- In the dark Intel-Night hallway, where RANSAC fails and falls back to constant motion (about 1.1 m mean error), RGB-D ICP keeps the mean error at 0.15 m (Sec. 4.2; Table 2)
- Two-Stage RGB-D ICP is only marginally less accurate than full RGB-D ICP but faster when RANSAC rarely fails (Sec. 4.2; Tables 2-3)
- Adding ICP point pairs lets SBA optimise sequences containing frames without enough RANSAC inliers, and SBA gives more consistent maps than TORO in visual comparisons (Secs. 3.2.3, 4.3; Figs. 5-7)
- Surfel map reduces 95 frames of about 250,000 points each from roughly 23,750,000 points to 730,000 surfels, a factor of 32, and is smoother than the point cloud (Sec. 4.4; Figs. 9, 11)
- Maps of two large indoor loops (71 m and 114 m) are consistent with a 2D laser map and an architectural floor plan (Sec. 4.1; Fig. 4)
限制
- RGB-D cameras give depth only up to a limited distance (typically less than 5 m), noisy depth (about 3 cm at 3 m) and a narrow field of view of about 60 deg (Sec. 1)
- Motion is estimated from only two consecutive frames; the authors suggest windowed bundle adjustment (Sec. 5)
- Loop closure by frame-to-frame visual matching detects obvious loops but does not add all constraints needed for fully consistent maps (Sec. 5)
- SBA is notably slower than TORO and its time grows roughly linearly with map size, so it will not scale to arbitrarily large maps without windowing or less frequent runs (Sec. 4.3; Fig. 8)
- Surfel maps are built as a post-process because pose updates from global optimisation require rebuilding them; surfel generation takes about 3 s per frame (Secs. 3.3, 4.4)
- Ground truth is limited to 16 inter-marker distances of 3 to 5.5 m; global consistency is judged visually against a 2D laser map and a floor plan, and the authors call for more extensive ground-truth experiments (Secs. 4.1, 4.2, 5)
- RANSAC is unreliable when features are out of depth range or concentrated in a small image region, for example in badly lit halls (Sec. 4.5)
營建工程相關證據
論文未在營建工地測試；兩個測試環境為辦公大樓室內大迴圈（Intel Labs Seattle 與華盛頓大學 Allen Center），全域一致性只以 2D 雷射地圖與建築平面圖疊合目視判斷，定量驗證限於 16 個標記間距離。此系統是以低成本 RGB-D 相機輸出室內稠密彩色點雲與面元地圖的早期範例，但作者所述量測範圍小於 5 m、3 m 處約 3 cm 雜訊與約 60 度視野，限制其直接用於施工量測或竣工點雲的精度要求（推論）。
原文驗證環境：受控實驗、獨立參考量測
報告的性能數據
性能數據仍在分批查證，目前尚未收錄此方法的報告值。
來源
Henry et al., 2012
（2012）RGB-D mapping: Using Kinect-style depth cameras for dense 3D modeling of indoor environmentsThe International Journal of Robotics Research, 31(5):647-663
同儕審查已出版已讀全文經典查證後修正
相關版本
- 會議版：RGB-D Mapping: Using Depth Cameras for Dense 3D Modeling of Indoor Environments (ISER 2010; Experimental Robotics, Springer Tracts in Advanced Robotics, 2014, pp. 477-491); the IJRR paper states it extends this version (Sec. 1); not read 10.1007/978-3-642-28572-1_33
- 預印本：Author preprint of the IJRR article (28 pages, PDF created 2011-12-05), formerly at homes.cs.washington.edu, retrieved from the Wayback Machine http://web.archive.org/web/2016id_/http://homes.cs.washington.edu/~peter/papers/3d-mapping-ijrr-12-preprint.pdf