RGBDSLAMv2
RGBDSLAMv2 只用 RGB-D 相機建立三維地圖。前端從彩色影像擷取 SIFT、SURF 或 ORB 特徵，以深度影像取得三維位置，再用 RANSAC 估計影格間的剛體轉換；候選影格包含前幾張影格、位姿圖上測地鄰域的抽樣以及關鍵影格，用來尋找迴圈閉合。作者提出以光束模型檢查深度影像間的自由空間衝突（EMM），剔除不可信的轉換，後端以 g2o 最佳化位姿圖並刪除誤差過大的邊。最後依軌跡把量測投影成點雲，或以 OctoMap 產生佔據體素地圖。
本頁內容
RGBDSLAMv2 builds 3-D maps from an RGB-D camera alone using keypoint matching with RANSAC, geodesic-neighbourhood and keyframe loop-closure candidates, a beam-based environment measurement model to reject wrong transforms, and g2o pose-graph optimisation with edge pruning, exporting point clouds or OctoMaps.
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 (structured light: Microsoft Kinect, Asus Xtion Pro Live) |
|---|---|
| 原文測試平台 | TUM RGB-D fr1 and fr2 sequences (the paper does not state how the camera was carried in these sequences) |
| 狀態估計 | pose-graph SLAM: pairwise 6-DOF transforms from 3-D feature correspondences via RANSAC with least-squares motion estimation, optional two-frame g2o refinement; global g2o optimisation (CSparse offline, PCG suggested online) with pruning of edges whose error remains high after convergence (Secs. III-B, III-E, IV-C) |
| 資料關聯 | sparse visual keypoints (SIFT on GPU, SURF, ORB or Shi-Tomasi plus SURF) matched by nearest to second-nearest ratio (Euclidean, Hellinger or Hamming distance), RANSAC with Mahalanobis inlier test; transforms validated by a beam-based environment measurement model (EMM) on subsampled depth images (Secs. III-B, III-C, IV-D) |
| 時間表示 | discrete frames; frame-to-frame transforms between RGB-D images |
| 去畸變 | 不適用 |
| 迴圈閉合 | candidate frames from n immediate predecessors, k frames sampled from the geodesic neighbourhood in the pose graph and l frames sampled from keyframes; validated by RANSAC and the EMM (Sec. III-D) |
| 全域最佳化 | g2o pose-graph optimisation with Mahalanobis-based edge pruning (Secs. III-E, IV-C) |
| 地圖表示 | globally registered point cloud (optionally surfels) or OctoMap 3-D occupancy grid; paper recommends OctoMap for memory and free-space representation (Sec. III-F) |
| 先驗資訊 | none (no odometry or other sensors used, even where wheel odometry was available) |
| 可輸出幾何 | optimised camera trajectory plus point cloud created by projecting the original depth measurements, or a textured OctoMap voxel occupancy map (2 cm maps of 4.2 to 25 MB versus 2 to 5 GB for unfiltered point clouds) (Sec. III-F) |
| 計算需求 | Intel Core i7 3.40 GHz with nVidia GeForce GTX 570 (SIFT on GPU); offline processing of every frame at 5.04 to 15.2 Hz in Table I; median 13.0 Hz (9.1 to 16.4 Hz) on fr1 with SIFT (Fig. 8 caption); EMM 0.82 ms per bidirectional check (Sec. IV-D) |
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 |
|---|---|---|---|---|---|
| RGB-D 相機 | Microsoft Kinect | 資料集感測器 | TUM RGB-D benchmark | structured light; two Kinect units used in the benchmark | (Endres et al., 2014, Sec. IV-A) |
| RGB-D 相機 | Asus Xtion Pro Live | 資料集感測器 | TUM RGB-D benchmark | structured light | (Endres et al., 2014, Sec. IV-A) |
| 載具平台 | Pioneer 3 | 資料集感測器 | TUM RGB-D Robot SLAM sequences | Kinect mounted on the robot; wheel odometry available but not used | (Endres et al., 2014, Sec. IV-A) |
| 運算硬體 | Intel Core i7 | 執行運算平台 | 未標示 | 3.40 GHz; used for all experiments | (Endres et al., 2014, Secs. IV, IV-B) |
| 運算硬體 | nVidia GeForce GTX 570 | 執行運算平台 | 未標示 | graphics card used for all experiments; SIFT computed on the GPU (SiftGPU) | (Endres et al., 2014, Secs. IV, IV-B) |
| 其他 | high-precision motion capturing system (model not reported) | 參考或真值量測 | TUM RGB-D benchmark | synchronised ground-truth sensor trajectory | (Endres et al., 2014, Secs. II, IV-A) |
作者報告的優勢與限制
優勢
- ATE RMSE 0.026 m on fr1/desk, 0.087 m on fr1/room and 0.057 m on fr2/desk; better than the best results reported for Kintinuous except fr2/large no loop (Table I)
- The EMM substantially reduces error on the challenging Robot SLAM (Pioneer) sequences, especially together with edge pruning (Figs. 9-10; Sec. IV-D)
- Geodesic-neighbourhood sampling reduced the average error on Robot SLAM by 26% (Fig. 5 caption)
- SIFT on GPU gives the highest accuracy (median RMSE 0.04 m on fr1) (Sec. IV-B)
- OctoMap output at 2 cm is 4.2 to 25 MB versus 2 to 5 GB for unfiltered point clouds (Sec. III-F)
限制
- Repetitive structures (same chairs, windows, wallpaper, poles) cause systematic misassociation and bogus transforms that can distort the optimised graph (Secs. III-B, III-E)
- Structured-light sensors are generally not usable in direct sunlight and have short range; feature-poor stretches with no depth occur in large halls (Secs. II, IV-A)
- ORB or Shi-Tomasi plus SURF reach real-time rates but with an average error of about 15 cm on fr1 (Sec. IV-B)
- OctoMap raycasting is expensive (about 1 s per 100 000 points at 5 cm; about 25 s per frame at 5 mm) and cannot be updated efficiently after large loop closures, so the map must be recreated (Sec. III-F)
- The chosen metric (ATE) does not directly assess map quality (Sec. IV-A)
營建工程相關證據
論文未在營建工地測試；評估使用 TUM RGB-D 資料集（辦公室尺度場景與工業廠房內的 Pioneer 機器人序列）以及 MIT Stata Center 序列。系統可直接輸出配準後點雲或 OctoMap，且作者指出結構光感測器在日光下無法使用、重複結構易造成錯誤配對，這些限制對室內裝修或機電階段的 RGB-D 掃描有參考價值（推論）。
原文驗證環境：公開基準、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
Endres et al., 2014
（2014）3-D Mapping With an RGB-D CameraIEEE Transactions on Robotics, 30(1):177-187
DOI 10.1109/tro.2013.2279412程式碼
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：An evaluation of the RGB-D SLAM system (ICRA 2012; Endres, Hess, Engelhard, Sturm, Cremers, Burgard); reference [24] of the T-RO paper matches this Crossref record (title, authors, ICRA 2012, pp. 1691-1696) 10.1109/ICRA.2012.6225199
- 程式碼釋出：felixendres/rgbdslam_v2 (ROS package; README says it creates 3D point clouds or OctoMaps and cites this T-RO article) https://github.com/felixendres/rgbdslam_v2
程式碼：https://github.com/felixendres/rgbdslam_v2（授權：GPL-3.0 (COPYING file; README states GPL v.3)）。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。