DVO-SLAM
DVO-SLAM 以稠密方式對齊 RGB-D 影像，同時最小化所有像素的光度誤差與深度誤差，並以雙變量 t 分布自動調整兩項誤差的權重，降低離群值影響。系統採影格對關鍵影格追蹤，以位姿估計共變異數的熵比值決定何時建立新關鍵影格，並用同一熵比值驗證以距離搜尋得到的迴圈閉合候選。所有成功的匹配組成位姿圖，以 g2o 最佳化以修正累積漂移。
本頁內容
DVO-SLAM aligns RGB-D frames densely by minimising t-distribution-weighted photometric and depth errors, selects keyframes and validates metric loop-closure candidates with an entropy-ratio test on the motion covariance, and optimises the keyframe pose graph with g2o.
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 |
|---|---|
| 原文測試平台 | TUM RGB-D benchmark sequences (fr1, fr2/desk, fr3/office and the structure and texture sequences)、the paper does not state how the camera was carried |
| 狀態估計 | dense frame-to-keyframe alignment minimising a bivariate t-distributed photometric and depth error by iteratively reweighted Gauss-Newton on se(3), coarse-to-fine over three resolutions up to 320 x 240; covariance from the inverse approximate Hessian; keyframe pose graph optimised with g2o (Secs. III, IV-C) |
| 資料關聯 | direct: all pixels warped with the depth map (photometric residual plus depth residual equivalent to point-to-plane ICP with projective lookup); no feature matching (Sec. III-D) |
| 時間表示 | discrete RGB-D frames |
| 去畸變 | 不適用 |
| 迴圈閉合 | metric nearest-neighbour search for keyframes within a sphere around the current keyframe, validated by the entropy ratio at coarse and then higher resolution; extra loop search for every keyframe at the end of a sequence (Sec. IV-B, IV-C) |
| 全域最佳化 | g2o pose-graph optimisation over keyframe poses with edges weighted by the covariance of the relative motion estimates (Sec. IV-C) |
| 地圖表示 | pose graph of keyframes (RGB-D images); a point cloud can be formed from the optimised trajectory (Fig. 1 text; Sec. IV-C) |
| 先驗資訊 | none |
| 可輸出幾何 | optimised keyframe trajectory; the corrected trajectory allows a consistent point cloud model of the scene, while high-quality 3D model generation is left as future work (Secs. I, IV-C, VI) |
| 計算需求 | Intel Core i7-2600 at 3.40 GHz with 16 GB RAM; frame-to-keyframe tracking about 32 ms; loop detection and optimisation (map update) averaging 135 ms in a separate thread (Sec. V) |
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 | RGB-D camera of the TUM RGB-D benchmark (model not named in this paper) | 資料集感測器 | TUM RGB-D | 原文未報告 (the method processes the images coarse-to-fine at three resolutions up to 320 x 240 pixels) | (Kerl et al., 2013, Sec. V) |
| 運算硬體 | Intel Core i7-2600 | 執行運算平台 | 未標示 | 3.40GHz, 16GB RAM; visual odometry and SLAM run in separate threads | (Kerl et al., 2013, Sec. V) |
| 其他 | external motion capture system (model not reported) | 參考或真值量測 | TUM RGB-D | accurate ground-truth trajectory | (Kerl et al., 2013, Sec. V) |
作者報告的優勢與限制
優勢
- Combined RGB and depth odometry is best on scenes with only structure or only texture and generalises better across scene types (Table I)
- Keyframes reduce drift by 16% on average and pose-graph optimisation raises the improvement to 20% (Table II)
- Average ATE RMSE on the fr1 sequences (excluding fr1/floor) drops from 0.19 m (frame-to-frame) to 0.07 m (keyframes plus optimisation) (Sec. V)
- Best ATE on five of eight datasets where all systems report results, with average ATE 0.034 m versus 0.054 m (RGB-D SLAM), 0.043 m (MRSMap) and 0.297 m (KinFu) (Table III)
- The depth term stabilises estimates during sudden intensity changes from auto-exposure (Sec. V)
限制
- Incorrect loop closures caused high drift on fr1/floor and fr1/plant (v) (Sec. V; Table II)
- Combined odometry is slightly worse than RGB-only on scenes with both structure and texture (Table I)
- Loop-closure search is metric and designed for space-restricted indoor scenes with accurate odometry (Sec. IV-B)
- Dense 3D model generation from the optimised trajectory was left for future work (Sec. VI)
營建工程相關證據
論文未在營建工地測試；全部評估使用 TUM RGB-D 公開資料集（真值由外部動作捕捉系統提供；論文未描述各序列的場景與相機攜帶方式）。作者指出迴圈搜尋以距離為準，適用於空間受限的室內環境，並在低紋理或低結構場景表現較佳，這對室內裝修階段以 RGB-D 相機掃描白牆或空房有參考價值，但系統未輸出量化的點雲精度（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
Kerl et al., 2013
（2013）Dense visual SLAM for RGB-D cameras2013 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 2100-2106
DOI 10.1109/iros.2013.6696650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
- 前身方法：Robust odometry estimation for RGB-D cameras (Kerl, Sturm, Cremers, ICRA 2013), the dense visual odometry that this paper extends (cited as [7]) 10.1109/ICRA.2013.6631104
- 程式碼釋出：tum-vision/dvo_slam (dvo_core, dvo_ros, dvo_slam, dvo_benchmark) https://github.com/tum-vision/dvo_slam
程式碼：https://github.com/tum-vision/dvo_slam（授權：GPL-3.0 for dvo_core, dvo_ros, dvo_slam and dvo_benchmark; bundled sophus under MIT (README licence section; no LICENSE file in the repository root)）。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。