Underground RGB-D 3DGS SLAM
本研究提出只用低成本 RGB-D 相機的地下空間 3DGS SLAM。前處理以 MSRCR、側窗濾波與 HIS 色彩空間正規化 gamma 校正增強低照度影像，並以預訓練深度補全網路（非局部傳播架構，未在地下資料上微調）填補深度空洞；追蹤以渲染色彩與深度殘差最佳化位姿，並以混合歐氏距離（0.3 m 或 15 影格）與重疊度兩階段選取關鍵影格；建圖以不透明度門檻（τ0 = 0.15）與加權觀測次數低於 5 剔除高斯；偵測迴圈後以 Levenberg-Marquardt 求解位姿圖，並剛性更新高斯。資料由搭載 Kinect2 的 Autolabor-Pro1 四輪機器人在煤礦巷道、車庫與地下室共 9 個場景（38.4 至 101.3 m²）蒐集，另以 TUM RGB-D 三序列驗證。軌跡真值來自 LiDAR、IMU 與相機融合的 SLAM 流程，並非獨立測量；地圖只以新視角 PSNR、SSIM、LPIPS 評估，作者明言沒有稠密三維幾何真值。9 個場景中本方法 ATE 為 4.8 至 9.6 cm，有 6 個場景在稠密方法中最低或並列最低，但 ORB-SLAM2 在 4 個場景更低、2 個場景持平。
本頁內容
RGB-D 3DGS SLAM for underground spaces with fixed-parameter low-light enhancement, a pretrained depth-completion network, two-stage keyframe selection, opacity and observation-count Gaussian pruning, and LM pose-graph loop closure. Evaluated on nine self-collected coal-mine, garage and basement sequences (Kinect2 on an Autolabor-Pro1 robot) with ATE against a LiDAR-IMU-camera SLAM reference, plus TUM RGB-D; maps are assessed only by novel-view PSNR, SSIM and LPIPS.
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 (Kinect2, 512x424 pixels, 10 Hz, 70 x 60 deg field of view) |
|---|---|
| 原文測試平台 | wheeled mobile robot (Autolabor-Pro1, four-wheel drive) carrying a Kinect2 RGB-D camera |
| 狀態估計 | Per-frame gradient-based pose optimization against colour and depth rendered from the Gaussian map (pixels with silhouette S(p) > 0.99 and valid depth; weighted colour and depth residuals), then sliding-window Gaussian optimization with poses fixed |
| 資料關聯 | Direct photometric and geometric rendering residuals against enhanced RGB and completed depth; keyframes by hybrid Euclidean pose distance (0.3 m or 15 frames) then overlap ratio; loop candidates by cosine similarity and geometric overlap following GLC-SLAM |
| 時間表示 | discrete per-frame poses; initial pose from a motion model (type not specified) |
| 去畸變 | 不適用 |
| 迴圈閉合 | Loop frames detected by cosine similarity and geometric overlap ratio (strategy similar to GLC-SLAM [58]); relative loop poses added as pose-graph edges |
| 全域最佳化 | Keyframe pose graph with odometry and loop edges minimized by Levenberg-Marquardt on the Lie group; Gaussian means and covariances rigidly updated with each keyframe correction |
| 地圖表示 | 3D Gaussian ellipsoids (following GSORB-SLAM [56]) with isotropic scale regularization; pruning by an opacity threshold (tau0 = 0.15; the text words it as transparency below zero or above 0.15) and by a weighted observation count below 5 |
| 先驗資訊 | Depth-completion network with non-local spatial propagation, pretrained on unnamed public RGB-D datasets and not fine-tuned on underground data; fixed-parameter image enhancement (MSRCR, side window filtering, normalized gamma correction in HIS space) |
| 可輸出幾何 | 3D Gaussian map; no 3D geometric accuracy evaluated (authors state dense geometric ground truth is unavailable); map quality reported only by novel-view PSNR, SSIM and LPIPS on held-out non-keyframes |
| 計算需求 | Processing on a desktop (Intel i7-14700K, RTX 4090D, 64 GB DDR5, Ubuntu 18.04, Python and PyTorch); about 8.6 FPS; per-frame preprocessing 72.5 ms, tracking 36.2 ms, mapping 41.8 ms; peak GPU 10.2 GiB; robot carries an Autolabor-PC control console (Ryzen 3 3200G, 8 GB) |
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 | 原文未報告 (LiDAR in ground-truth fusion pipeline) | 參考或真值量測 | Underground_RGB-D (authors' field test dataset) | 原文未報告 | (Yan et al., 2026b, Sec. IV-A) |
| 慣性量測單元（IMU） | 原文未報告 (IMU in ground-truth fusion pipeline) | 參考或真值量測 | Underground_RGB-D (authors' field test dataset) | 原文未報告 | (Yan et al., 2026b, Sec. IV-A) |
| 相機 | 原文未報告 (camera in ground-truth fusion pipeline) | 參考或真值量測 | Underground_RGB-D (authors' field test dataset) | 原文未報告 | (Yan et al., 2026b, Sec. IV-A) |
| RGB-D 相機 | Kinect2歸入：Microsoft Kinect v2 | 方法輸入 | Underground_RGB-D (authors' field test dataset) | Sampling rate 10 Hz; resolution 512x424 pixels; horizontal 70 deg, vertical 60 deg | (Yan et al., 2026b, Table I; Sec. IV-A; Fig. 4) |
| 載具平台 | Autolabor-Pro1 | 方法輸入 | Underground_RGB-D (authors' field test dataset) | Mobile robot, four-wheel drive; displacement speed 0.5 to 1.5 m/s; angular velocity 0.56 rad/s | (Yan et al., 2026b, Table I; Fig. 4) |
| 運算硬體 | Autolabor-PC (control console) | 方法輸入 | Underground_RGB-D (authors' field test dataset) | CPU AMD Ryzen3 3200G, DDR4 8GB; listed in Table I as the control console of the data-collection platform, while all SLAM experiments ran on the desktop computer (Sec. IV-A) | (Yan et al., 2026b, Table I) |
| 運算硬體 | Custom desktop computer (Intel i7-14700K, NVIDIA GeForce RTX 4090D) | 執行運算平台 | 未標示 | CPU Intel i7-14700K; GPU NVIDIA GeForce RTX 4090D; DDR5 64GB; Ubuntu 18.04 | (Yan et al., 2026b, Table I; Sec. IV-A; Sec. IV-C) |
作者報告的優勢與限制
優勢
- 13.8% higher PSNR than the best baseline in the underground tests (abstract; Table IV)
- Lowest or tied-lowest ATE among the eight tested NeRF and 3DGS SLAM methods in 6 of 9 field sequences, e.g., 4.8 cm on the pedestrian walkway (Table III)
- About 8.6 FPS with 10.2 GiB peak GPU memory; authors state this is comparable to SplaTAM and MonoGS but slower than GS-ICP SLAM and RTG-SLAM (Sec. IV-C; Table VII)
- Removing enhancement, completion and loop closure raises ATE from 8.2 to 14.9 cm and lowers PSNR from 22.4 to 18.2 dB (Table VIII)
- Field dataset released at huggingface.co/datasets/MISPM/Underground_RGB-D (Sec. IV-A)
限制
- No 3D geometric accuracy is evaluated; authors state dense geometric ground truth is unavailable and assess maps only by novel-view rendering (Sec. IV-A; Sec. IV-B2)
- ATE reference comes from a multi-sensor (LiDAR, IMU, camera) SLAM pipeline whose sensor models and accuracy are not quantified; the ATE alignment method is not stated (Sec. IV-A)
- ORB-SLAM2 attains lower ATE than the proposed method in 4 of 9 field sequences and ties in 2; on TUM RGB-D ORB-SLAM2 is better on all three sequences (Tables III and VI)
- About 8.6 FPS does not meet higher frame-rate robotic needs; preprocessing (72.5 ms per frame) is the bottleneck (Sec. IV-C; Table V; Sec. V)
- Ablation text and Table VIII disagree (11.0 cm ATE paired with 22.2 dB in the text, but with 20.6 dB in the table), and the ablation scene is not identified (Sec. IV-E; Table VIII)
- Local artifacts in top views under severe illumination imbalance and repetitive textures (Sec. IV-B3)
- (inference) Depth-completed regions are network predictions and are not flagged in the map; no code link is given
營建工程相關證據
場域為煤礦巷道（回採巷道、運輸巷道、行人巷道、受限通道）、廢棄車庫與電動機車車庫，以及泵房、配電室與簡報室等地下室空間，皆為營運中或既有設施，非施工中工地，對隧道與地下設施數位孿生有參考價值。軌跡以多感測器 SLAM 結果為參考，地圖只評估渲染品質而缺少三維幾何精度；深度補全區域屬學習推定幾何，論文未另行標記（推論）。
原文驗證環境：地下或隧道、已完工建築、公開基準
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 4 個比較組，合計 16 筆紀錄。
Yan et al., 2026b · Table III 本方法 9 筆
指標ATE [cm]
表格設定（擷取紀錄原文）：ATE RMSE on the authors' nine underground RGB-D field sequences (Kinect2 on a mobile robot); trajectory reference from a multi-sensor (LiDAR, IMU, camera) fusion SLAM pipeline, not an independent survey; all methods rerun with official code on the same PC (Sec. IV-A); values transcribed from the table image (Yan et al., 2026b, Table III)
ATE [cm]，Underground_RGB-D (authors' field test dataset) · Extraction roadway (E-r)
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Yan et al., 2026b 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Yan et al., 2026b, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| NICE-SLAM | 18.9 cm | (Yan et al., 2026b, Table III) |
| Co-SLAM | 11 cm | (Yan et al., 2026b, Table III) |
| ESLAM | 9.9 cm | (Yan et al., 2026b, Table III) |
| SplaTAM | 12.6 cm | (Yan et al., 2026b, Table III) |
| MonoGS | 15.2 cm | (Yan et al., 2026b, Table III) |
| GS-ICP SLAM | 9.5 cm | (Yan et al., 2026b, Table III) |
| RTG-SLAM | 9.8 cm | (Yan et al., 2026b, Table III) |
| Ours本方法原文提出 | 9.1 cm | (Yan et al., 2026b, Table III) |
| ElasticFusion | 13.2 cm | (Yan et al., 2026b, Table III) |
| ORB-SLAM2 | 8.9 cm | (Yan et al., 2026b, Table III) |
Yan et al., 2026b · Table V 本方法 3 筆
資料集與序列Underground_RGB-D (authors' field test dataset) · average over sequences
表格設定（擷取紀錄原文）：Average time per frame for each stage, averaged over multiple underground sequences (Yan et al., 2026b, Table V)
Average duration, Preprocessing (image enhancement and depth completion)，Underground_RGB-D (authors' field test dataset) · average over sequences
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 Yan et al., 2026b 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 |
|---|---|---|
| Ours本方法原文提出硬體：Intel i7-14700K, NVIDIA GeForce RTX 4090D, 64 GB DDR5, Ubuntu 18.04 (Python, PyTorch) | 72.5 ms | (Yan et al., 2026b, Table V) |
Yan et al., 2026b · Table VI 本方法 3 筆
指標ATE [cm]
表格設定（擷取紀錄原文）：ATE RMSE on three TUM RGB-D sequences; ElasticFusion and ORB-SLAM2 values are identical to those printed in SplaTAM Table 1 (from Point-SLAM), suggesting reuse of published values although Sec. IV-A states all comparisons were reproduced (inference) (Yan et al., 2026b, Table VI)
ATE [cm]，TUM RGB-D · fr1/desk
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Yan et al., 2026b 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Yan et al., 2026b, Table VI)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| NICE-SLAM | 3.2 cm | (Yan et al., 2026b, Table VI) |
| Co-SLAM | 2.7 cm | (Yan et al., 2026b, Table VI) |
| ESLAM | 2.8 cm | (Yan et al., 2026b, Table VI) |
| SplaTAM | 3.3 cm | (Yan et al., 2026b, Table VI) |
| MonoGS | 3.2 cm | (Yan et al., 2026b, Table VI) |
| GS-ICP-SLAM | 2.7 cm | (Yan et al., 2026b, Table VI) |
| RTG-SLAM | 2.1 cm | (Yan et al., 2026b, Table VI) |
| Ours本方法原文提出 | 2.1 cm | (Yan et al., 2026b, Table VI) |
| ElasticFusion | 2.53 cm | (Yan et al., 2026b, Table VI) |
| ORB-SLAM2 | 1.6 cm | (Yan et al., 2026b, Table VI) |
Yan et al., 2026b · Text Sec. IV-C 本方法 1 筆
指標average frame rate
資料集與序列Underground_RGB-D (authors' field test dataset) · average over sequences
表格設定（擷取紀錄原文）：Average frame rate over multiple underground sequences (Yan et al., 2026b, Text Sec. IV-C)
average frame rate，Underground_RGB-D (authors' field test dataset) · average over sequences
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 Yan et al., 2026b 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 |
|---|---|---|
| Ours本方法原文提出硬體：Intel i7-14700K, NVIDIA GeForce RTX 4090D, 64 GB DDR5, Ubuntu 18.04 (Python, PyTorch) | 8.6 FPS | (Yan et al., 2026b, Sec. IV-C) |
來源
Yan et al., 2026b
（2026）RGB-D Perception-Enhanced 3D Gaussian Splatting SLAM: A Robust Framework for Mapping Underground SpacesIEEE Transactions on Visualization and Computer Graphics, 32(7), 6695-6711
同儕審查已出版已讀全文近十年查證後修正