OGM-enhanced visual SLAM (Xu et al. 2019)
作者以 ORB-SLAM2 的 RGB-D 模式為基礎，建立可用於室內即時定位系統（RTLS）的視覺 SLAM。除了 ORB-SLAM2 原有的稀疏特徵地圖外，系統把 Kinect 點雲在指定高度範圍內切成虛擬雷射掃描，搭配關鍵影格位姿投影成二維佔據格地圖；為避免早期位姿誤差污染地圖，只用關鍵影格建圖，並在固定數量關鍵影格後或迴圈閉合修正後，以最佳化過的全部歷史關鍵影格重建佔據格地圖。系統可在 SLAM 模式與只定位模式間切換，佔據格地圖則讓使用者互動、A* 路徑規劃、依位置觸發的資料蒐集與局部點雲更新成為可能。定位精度以地下室走廊 15 個 AprilTag 標記的位置與間距評估。
本頁內容
ORB-SLAM2 RGB-D extended with a 2D occupancy grid built from keyframe poses and virtual laser scans cut from the Kinect point cloud (rebuilt after loop correction), giving an indoor real-time locating system with SLAM and localization modes; accuracy assessed with 15 floor markers in an 80 m basement corridor loop.
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 (Microsoft Kinect, via ROS openni_launch; depth registered to RGB)、virtual 2D laser scan cut from the Kinect point cloud (pointcloud_to_laserscan) |
|---|---|
| 原文測試平台 | robotic wheelchair carrying a Kinect (corridor evaluation)、robotic powerchair and TurtleBot (application demonstrations)、laboratory room tests (carrier not specified) |
| 狀態估計 | ORB-SLAM2 RGB-D (ORB2 RGBD) for 3D camera poses in SLAM or localization-only mode; a separate occupancy-grid mapping node projects keyframe poses to 2D and integrates virtual laser scans with log-odds updates (lfree 5, locc 15), rebuilding the OGM from all optimized historical keyframes periodically and after loop correction (Sec. 4.1) |
| 資料關聯 | ORB features and bag-of-words relocalization of ORB-SLAM2; OGM cells updated by Bresenham ray tracing of virtual scan beams (Sec. 4.1.4, 4.2) |
| 時間表示 | discrete keyframes; RGB, depth and virtual scan synchronized into one custom ROS message |
| 去畸變 | 不適用 (RGB-D camera) |
| 迴圈閉合 | ORB-SLAM2 loop detection and correction; OGM rebuilt with the corrected keyframe poses (Sec. 4.1.3) |
| 全域最佳化 | ORB-SLAM2 local mapping, loop closing and pose-graph correction; saved feature maps serialized with the scans to allow incremental map extension across sessions |
| 地圖表示 | sparse ORB feature map plus a 2D occupancy grid map (ROS OccupancyGrid); colored RGB-D point cloud built incrementally for the application example |
| 先驗資訊 | none (maps built by the system itself in SLAM mode, then reused in localization mode) |
| 可輸出幾何 | 2D occupancy grid with real-time 2D camera pose and virtual scan overlay; incremental colored point cloud from RGB-D frames (Sec. 6.3) |
| 計算需求 | Localization mode about 17.1 FPS average (12.8 to 24.4) at 640 x 480 on a laptop with Intel Core i7-4940MX 3.1 GHz (Sec. 4.2) |
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 | Kinect | 方法輸入 | 未標示 | RGB and depth 640 x 480; openni_launch driver with depth registration to RGB | (Xu et al., 2019, Sec. 3, Sec. 4.1.1, Sec. 4.2) |
| 載具平台 | robotic wheelchair equipped with a Kinect | 方法輸入 | 未標示 | moved for five complete loops along an 80 m basement corridor | (Xu et al., 2019, Sec. 5.2.2) |
| 載具平台 | robotic powerchair | 方法輸入 | 未標示 | localized in real time on the OGM for A* path planning and navigation demonstrations | (Xu et al., 2019, Sec. 6.1, Fig. 13) |
| 載具平台 | TurtleBot | 方法輸入 | 未標示 | data collection platform for the geo-tagged environmental data collection example | (Xu et al., 2019, Sec. 6.2, Fig. 14) |
| 運算硬體 | laptop with Intel Core i7-4940MX CPU @ 3.1GHz | 執行運算平台 | 未標示 | localization speed evaluation | (Xu et al., 2019, Sec. 4.2) |
| 其他 | AprilTag markers (15) | 參考或真值量測 | 未標示 | fixed on the corridor floor; poses detected by the AprilTag algorithm | (Xu et al., 2019, Sec. 5.1-5.2) |
作者報告的優勢與限制
優勢
- Marker position RMSE 0.039 to 0.186 m and marker distance RMSE 0.018 to 0.235 m in a low-feature basement corridor loop (Tables 1-2)
- Repeated measurement of a fixed marker from six locations gave standard deviations of 0.002, 0.002 and 0.003 m (Sec. 5.2.1)
- Localization mode runs at about 17.1 FPS on a laptop, above the 1 Hz real-time threshold used for construction equipment tracking (Sec. 4.2)
- No environment instrumentation or prior floor plan required; switching between SLAM and localization modes allows incremental map extension (Sec. 3-4)
限制
- Largest errors occurred where the wheelchair passed close to low-feature walls (markers #7 and #9) (Sec. 5.2.2)
- Feature-based ORB-SLAM2 struggles with severe illumination change, repetitive features, large viewpoint differences and non-smooth rotation in corridors (Sec. 7)
- Only indoor RGB-D SLAM with a Kinect was evaluated (Sec. 3)
- Virtual laser scans are of lower quality than real laser scanner data and are used only for mapping, not localization (Sec. 4.1.2)
營建工程相關證據
作者把系統定位為營建與設施管理機器人的室內即時定位方案，並稱測試場地為兩種典型室內營建環境，但實際描述的是大學實驗室房間與建築地下室走廊，文中沒有說明是否處於施工階段；走廊評估使用 15 個地面標記，世界座標以同一組標記求得的轉換對齊後才比較（Sec. 5、Sec. 5.2.2）。
原文驗證環境：已完工建築、受控實驗
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 36 筆紀錄。
Xu et al., 2019 · Table 1 本方法 15 筆
指標marker position RMSE (2D)
表格設定（擷取紀錄原文）：Localization mode on the building-scale maps; 15 AprilTag markers on the floor of an 80 m basement corridor loop, five loops with a robotic wheelchair; estimated 2D marker positions compared with measured true positions after aligning the system world frame to the marker network frame with a 3D transformation from the same 15 marker pairs (Xu et al., 2019, Table 1)
marker position RMSE (2D)，authors' basement corridor data · marker #1 (76 measurements)
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 Xu 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 OGM-enhanced ORB2 RGB-D RTLS本方法原文提出 | 0.098 m | (Xu et al., 2019, Table 1) |
Xu et al., 2019 · Table 2 本方法 15 筆
指標marker distance RMSE
表格設定（擷取紀錄原文）：Distance between adjacent markers estimated by the system versus measured true distance, same corridor experiment (Xu et al., 2019, Table 2)
marker distance RMSE，authors' basement corridor data · markers #1-#2 (true distance 3.05 m)
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 Xu 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 OGM-enhanced ORB2 RGB-D RTLS本方法原文提出 | 0.018 m | (Xu et al., 2019, Table 2) |
Xu et al., 2019 · Text Sec.4.2 本方法 3 筆
資料集與序列authors' laboratory data · localization mode
表格設定（擷取紀錄原文）：Localization-mode update rate of the 2D pose on the OGM as reported by ROS rostopic, 640 x 480 RGB and registered depth images (Xu et al., 2019, Text Sec.4.2)
mean localization speed (about)，authors' laboratory data · localization mode
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 Xu 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 OGM-enhanced ORB2 RGB-D RTLS (localization mode)本方法原文提出硬體：laptop Intel Core i7-4940MX CPU @ 3.1 GHz | 17.1 FPS有附註註記（擷取紀錄）：other: approximate ('~') value stated in text | (Xu et al., 2019, Sec. 4.2) |
Xu et al., 2019 · Text Sec.5.2.1 本方法 3 筆
資料集與序列authors' laboratory room data · single marker, 600 measurements
表格設定（擷取紀錄原文）：Repeatability test: one fixed marker measured 100 times from each of six camera locations (600 measurements) in the laboratory-scale map; accuracy not evaluated because the world frame is arbitrary (Xu et al., 2019, Text Sec.5.2.1)
standard deviation of marker x position，authors' laboratory room data · single marker, 600 measurements
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 Xu 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 OGM-enhanced ORB2 RGB-D RTLS本方法原文提出 | 0.002 m | (Xu et al., 2019, Sec. 5.2.1) |
來源
Xu et al., 2019
（2019）An Occupancy Grid Mapping enhanced visual SLAM for real-time locating applications in indoor GPS-denied environmentsAutomation in Construction, 104:230-245
DOI 10.1016/j.autcon.2019.04.011
同儕審查已出版已讀全文近十年查證後修正