DEMO
DEMO 以單眼相機為主，從 RGB-D 相機或 LiDAR 取得深度：先用估測的運動把深度點登錄成局部深度地圖並存進以兩個角度座標建立的 2D KD 樹，再以最近三點構成的小平面內插特徵深度；沒有深度的特徵改用前幾影格的運動三角化，仍無法取得時也保留並以較弱的約束參與求解。逐影格運動用具強健權重的 Levenberg-Marquardt 求解，另以 iSAM 對 8 張影像做低頻光束法平差，最後整合成高頻輸出。
本頁內容
Monocular visual odometry that attaches depth from an RGB-D camera or a LiDAR through a motion-registered depth map, uses features both with and without depth in a robust frame-to-frame solve, and refines motion with low-rate windowed bundle adjustment (iSAM).
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 | monocular camera、depth from an RGB-D camera (Xtion Pro Live) or from a 3D LiDAR (rotating Hokuyo UTM-30LX; Velodyne on KITTI) |
|---|---|
| 原文測試平台 | handheld or mobile sensor rigs (author-collected room, lobby, road and lawn tests)、KITTI vehicle |
| 狀態估計 | frame-to-frame motion by Levenberg-Marquardt robust fitting with bisquare weights, using features with depth (two equations each) and without depth (one equation each); sliding bundle adjustment with iSAM on 8 images (one of every five of 40 frames) at about 0.25 to 1.0 Hz; transform integration combines the high-rate and low-rate estimates (Secs. V-VI) |
| 資料關聯 | Harris corners tracked by KLT; depth map registered with the estimated motion, downsampled by angular interval and stored in a 2D KD-tree over two angular coordinates; feature depth interpolated from three nearest depth points (planar patch) or, if unavailable, triangulated from previous motion (Sec. V) |
| 時間表示 | discrete image frames (30 Hz in the author tests; 10 Hz on KITTI); depth points are registered into the depth map using the estimated camera motion |
| 去畸變 | not described; LiDAR points are accumulated into the depth map using the estimated motion |
| 迴圈閉合 | no |
| 全域最佳化 | none beyond the windowed bundle adjustment |
| 地圖表示 | local registered depth map (point cloud) around the camera used only for depth association |
| 先驗資訊 | pre-calibrated camera intrinsics and camera-depth sensor extrinsics (Sec. III) |
| 可輸出幾何 | camera trajectory; registered point clouds of the depth sensor (Fig. 7 shows maps built on KITTI) |
| 計算需求 | laptop with 2.5 GHz cores and 6 GB memory, about three cores used; real-time at the camera rate (Sec. VII) |
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 | UTM-30LX (motor-rotated)歸入：Hokuyo UTM-30LX | 方法輸入 | DEMO author-collected tests | 180 deg FoV, 0.25 deg resolution, 40 lines/s; rotated by a motor for 3D scanning | (Zhang et al., 2014, Sec. III; Fig. 2(b)) |
| LiDAR | KITTI 360 deg Velodyne laser scanner | 方法輸入 | KITTI odometry | 10 Hz logging | (Zhang et al., 2014, Sec. VII) |
| GNSS 接收器 | KITTI high-accuracy GPS/INS歸入：KITTI high accuracy GPS/INS | 參考或真值量測 | KITTI odometry | ground truth for KITTI | (Zhang et al., 2014, Sec. VII) |
| 相機 | custom-built camera | 方法輸入 | DEMO author-collected tests | up to 60 Hz, 744 x 480, 83 deg horizontal FoV | (Zhang et al., 2014, Sec. III; Fig. 2(b)) |
| 相機 | KITTI left monochrome camera | 方法輸入 | KITTI odometry | 10 Hz logging | (Zhang et al., 2014, Sec. VII) |
| RGB-D 相機 | Xtion Pro Live | 方法輸入 | DEMO author-collected tests | RGB and depth at 30 Hz, 640 x 480, 58 deg horizontal FoV | (Zhang et al., 2014, Sec. III; Fig. 2(a)) |
| 運算硬體 | laptop | 執行運算平台 | 未標示 | 2.5 GHz cores, 6 GB memory, about three cores used | (Zhang et al., 2014, Sec. VII) |
作者報告的優勢與限制
優勢
- Relative error at trajectory end 1.53 to 3.72 % (RGB-D) and 0.79 to 2.06 % (LiDAR) in four author tests, while Fovis and DVO degrade strongly as depth coverage shrinks in open scenes (Table I)
- KITTI training sequences 00-10 mean relative position error 0.93 to 1.87 % (Table II)
- Rated second on the KITTI odometry benchmark irrespective of sensing modality and first among visual odometry methods at the time of writing (Abstract)
- Bundle adjustment reduced mean error by 0.3 to 0.7 % on KITTI (Sec. VII)
- Still uses features without depth, so it keeps working when depth covers only part of the image (Sec. V; Fig. 6)
限制
- Feature tracking (Harris plus KLT) is unreliable in homogeneously coloured corridors (Sec. VIII)
- Only RGB-D and LiDAR depth were tested; stereo depth is future work (Sec. VIII)
- No loop closure; drift accumulates with distance
- Bundle adjustment is less effective on highway scenes, partly due to lower feature quality (Sec. VII)
營建工程相關證據
DEMO 是早期把 LiDAR 或 RGB-D 深度接到單眼視覺里程計的代表方法，後續視覺 LiDAR 里程計（如 SDV-LOAM）仍以它作為比較基準。論文測試包含室內房間、大廳與戶外道路、草地，沒有施工現場或點雲精度評估；對工地而言，其價值主要在說明深度覆蓋率下降時如何維持追蹤（推論）。
原文驗證環境：公開基準、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 3 個比較組，合計 32 筆紀錄。
Yuan et al., 2023a · Table V 本方法 13 筆
指標Relative translational error (RTE)
表格設定（擷取紀錄原文）：KITTI odometry; relative translational error (%) of LiDAR-assisted depth-enhanced visual odometry; baseline values from the original publications; LIMO* uses semantic information; '-' cells not stored (Yuan et al., 2023a, Table V)
Relative translational error (RTE)，KITTI odometry · 00
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 Yuan et al., 2023a 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(Yuan et al., 2023a, Table V)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| DEMO本方法 | 1.05% | (Yuan et al., 2023a, Table V; Sec. VII) |
| LIMO* | 1.12% | (Yuan et al., 2023a, Table V; Sec. VII) |
| Huang et al. | 0.99% | (Yuan et al., 2023a, Table V; Sec. VII) |
| DVL-SLAM | 0.93% | (Yuan et al., 2023a, Table V; Sec. VII) |
| Our VO module原文提出 | 0.67% | (Yuan et al., 2023a, Table V; Sec. VII) |
Zhang et al., 2014 · Table II 本方法 11 筆
指標Mean relative position error
表格設定（擷取紀錄原文）：KITTI odometry training sequences 00-10; left monochrome camera plus Velodyne; mean relative position error as a percentage (KITTI protocol) (Zhang et al., 2014, Table II)
Mean relative position error，KITTI odometry · 00 (3714 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 Zhang et al., 2014 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 |
|---|---|---|
| DEMO (camera + LiDAR)本方法原文提出 | 1.05% | (Zhang et al., 2014, Table II; Sec. VII) |
Zhang et al., 2014 · Table I 本方法 8 筆
指標Relative position error (% of distance traveled)
表格設定（擷取紀錄原文）：Author-collected tests with the Xtion RGB-D camera and with the custom camera plus rotating Hokuyo LiDAR; the camera starts and stops at the same position and the gap between the trajectory ends divided by trajectory length is the relative position error (3D coordinates); Fovis and DVO use RGB-D input (Zhang et al., 2014, Table I)
Relative position error (% of distance traveled)，DEMO author-collected tests · Room (16 m)
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 Zhang et al., 2014 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(Zhang et al., 2014, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Fovis | 2.72% | (Zhang et al., 2014, Table I; Sec. VII) |
| DVO | 1.87% | (Zhang et al., 2014, Table I; Sec. VII) |
| Our VO (RGB-D)本方法原文提出 | 2.14% | (Zhang et al., 2014, Table I; Sec. VII) |
| Our VO (Lidar)本方法原文提出 | 2.06% | (Zhang et al., 2014, Table I; Sec. VII) |
來源
Zhang et al., 2014
（2014）Real-time depth enhanced monocular odometry2014 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 4973-4980
DOI 10.1109/iros.2014.6943269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
- successor：A real-time method for depth enhanced visual odometry (Autonomous Robots 41(1):31-43, 2017; online 2015-12-12; not read) https://doi.org/10.1007/s10514-015-9525-1
- 程式碼釋出：demo_rgbd and demo_lidar ROS packages named in the paper (pages behind an anti-bot challenge; availability not verified) http://wiki.ros.org/demo_rgbd
程式碼：http://wiki.ros.org/demo_rgbd（授權：not verified (ROS wiki pages behind an anti-bot challenge; not bypassed)）。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。