V-LOAM
V-LOAM 以單眼相機搭配掃描式 3D 光達（由馬達帶動的 Hokuyo 2D 雷射掃描儀），分成兩個依序運作的階段：視覺里程計以影像速率（60 Hz）估計相鄰影格間的運動，特徵點的深度取自光達深度圖或三角化，沒有深度的特徵也納入求解；光達里程計每次掃描（約 1 秒）執行一次，先以線性運動模型做掃描對掃描精修，消除視覺漂移造成的點雲畸變，再以邊緣與平面特徵做掃描對地圖配準並累積地圖。最後整合低頻光達位姿與高頻視覺運動，以影像速率輸出位姿。方法與硬體都沒有使用 IMU，也刻意不做迴圈閉合。
本頁內容
Visual odometry provides high-rate, low-fidelity motion for registering scanning-LiDAR points, which LiDAR scan-matching odometry then refines.
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 (uEye monochrome at 60 Hz; wide-angle lens 76 deg or fisheye lens 185 deg horizontal FoV)、3D lidar built from a Hokuyo UTM-30LX 2D laser scanner rotated back-and-forth by a motor with encoder (1 s sweep)、KITTI configuration: single camera and Velodyne lidar、no IMU in the method or hardware |
|---|---|
| 原文測試平台 | vehicle (KITTI odometry benchmark)、handheld (custom camera-lidar sensor carried by a person; about 0.7 m/s in Tests 1 and 2) |
| 狀態估計 | two sequential stages: frame-to-frame visual odometry solved by Levenberg-Marquardt in a robust-fitting framework, using features with depth from the lidar depthmap, depth from triangulation, or no depth; lidar odometry once per sweep with sweep-to-sweep refinement (linear drift model) and then sweep-to-map registration; transforms from both stages integrated into poses at image rate (Secs. IV to VI) |
| 資料關聯 | up to 300 Harris corners tracked by KLT over 5x6 image subregions; feature depth interpolated from the three nearest depthmap points found in a 2D KD-tree on angular coordinates; lidar edge and planar points selected by local curvature and matched to edge lines and planar patches with 3D KD-trees (sweep-to-sweep) or by eigenvalue analysis of local map clusters (sweep-to-map, ICP-style) (Secs. V to VII) |
| 時間表示 | discrete frame-to-frame motion at the 60 Hz image rate; visual odometry drift modeled as linear (constant-velocity) motion within each 1 s sweep; lidar odometry at 1 Hz (Secs. I, IV, VI) |
| 去畸變 | points registered with the visual odometry motion; residual distortion from visual drift removed by the linear-motion model in the sweep-to-sweep refinement (Sec. VI, Fig. 4) |
| 迴圈閉合 | none; the authors intentionally omit loop closure to focus on odometry (Sec. I) |
| 全域最佳化 | none |
| 地圖表示 | incrementally built map point cloud: each distortion-free sweep is matched to the existing map cloud, with correspondences found by eigenvalue analysis of local point clusters, and then merged into it; edge and planar points of the previous sweep are kept in two 3D KD-trees for sweep-to-sweep matching (Sec. VI, Fig. 6) |
| 先驗資訊 | none in the reported experiments; the authors state the method can be configured for localization only if a prior map is available (Sec. I) |
| 可輸出幾何 | registered 3D point cloud maps (Figs. 9 to 14) and 6-DoF poses at the image frame rate; export format 原文未報告 |
| 計算需求 | real-time on a laptop with 2.5 GHz quad cores under Linux; about 2.5 cores in total, 2 for visual odometry and 0.5 for lidar odometry (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 | Hokuyo UTM-30LX | 方法輸入 | 未標示 | 2D laser scanner, 180 deg FoV, 0.25 deg resolution, 40 lines/s; motor-actuated to form a 3D lidar | (Zhang & Singh, 2015, Sec. VII, Fig. 8) |
| LiDAR | Velodyne lidar of the KITTI setup | 資料集感測器 | KITTI odometry benchmark | 原文未報告 | (Zhang & Singh, 2015, Sec. VII) |
| 相機 | uEye monochrome camera | 方法輸入 | 未標示 | 60 Hz, 752 x 480 px, wide-angle lens with 76 deg horizontal FoV | (Zhang & Singh, 2015, Sec. VII, Sec. VII-A, Fig. 8) |
| 相機 | second camera mounted underneath the original uEye camera, with fisheye lens | 方法輸入 | 未標示 | same configuration as the original camera except resolution 640 x 480 px; fisheye lens with 185 deg horizontal FoV; model not named separately | (Zhang & Singh, 2015, Sec. VII-A) |
| 相機 | single camera of the KITTI setup | 資料集感測器 | KITTI odometry benchmark | 原文未報告 | (Zhang & Singh, 2015, Sec. VII) |
| 載具平台 | handheld custom camera-lidar sensor | 方法輸入 | 未標示 | carried by a person walking at about 0.7 m/s in the accuracy tests | (Zhang & Singh, 2015, Sec. VII-A) |
| 運算硬體 | laptop with 2.5 GHz quad cores | 執行運算平台 | 未標示 | Linux; method uses about 2.5 cores | (Zhang & Singh, 2015, Sec. VII) |
| 其他 | motor and encoder actuating the scanner | 方法輸入 | 未標示 | rotates back-and-forth at 180 deg/s between -90 and 90 deg; encoder resolution 0.25 deg; one sweep lasts 1 s | (Zhang & Singh, 2015, Sec. IV, Sec. VII, Fig. 8) |
作者報告的優勢與限制
優勢
- Ranked first on the KITTI odometry benchmark at the time, 0.75% relative position drift (abstract, Sec. VIII)
- Relative position error 0.31% to 0.73% on own handheld tests of 47 m to 538 m; wide-angle and fisheye setups reach the same level after lidar refinement (Table I)
- With a fisheye camera, fast trials finished at 1.3% (staircase, up to about 170 deg/s) and 0.39% (corridor, about 2.6 m/s) where the wide-angle setup failed (Table II, Sec. VII-B)
- Tolerates 2 s light outages by constant-velocity prediction corrected by lidar odometry (Sec. VII-B, Fig. 14)
限制
- Unsuitable for continuous darkness; the authors recommend lidar-only LOAM there (Sec. VII-B)
- Fast motion blurs the point cloud and bends walls (Figs. 12 to 13)
- The wide-angle camera loses feature tracking in fast turns, so motion estimation fails (Table II)
- Scan matching fails in degenerate scenes dominated by planar areas (Sec. I)
- (inference) Accuracy is measured from loop gaps, manually matched points or assumed flat walls, without an independent reference instrument (Sec. VII)
營建工程相關證據
未在施工工地測試。自建資料為手持感測器在建物室內、穿越建物的室內外路徑（538 m）、含七個 180 度轉彎的樓梯間與走廊的測試，另有關燈造成光照劇變的室內測試（Sec. VII）；並以 KITTI 車載資料驗證。精度以迴圈缺口、人工對應點或起終點差估算，沒有獨立參考量測。與營建的關聯僅在於樓梯間與走廊等既有建物情境。
原文驗證環境：公開基準、已完工建築
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 6 個比較組，合計 28 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 2 組列在最後，並連到性能比較頁。
Zhang & Singh, 2015 · Table I 本方法 16 筆
指標Relative position error
表格設定（擷取紀錄原文）：Relative position error as a fraction of distance travelled, based on 3D coordinates; no independent reference instrument (Zhang & Singh, 2015, Table I)
Relative position error，authors' custom camera-lidar sensor data · Test 1 (Loop 1), 49 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 & Singh, 2015 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 & Singh, 2015, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| W-V (wide-angle camera, visual odometry only)本方法 | 1.1 | (Zhang & Singh, 2015, Table I, Sec. VII-A) |
| F-V (fisheye camera, visual odometry only)本方法 | 1.8 | (Zhang & Singh, 2015, Table I, Sec. VII-A) |
| W-VL (wide-angle camera, V-LOAM visual plus lidar odometry)本方法原文提出 | 0.31 | (Zhang & Singh, 2015, Table I, Sec. VII-A) |
| F-VL (fisheye camera, V-LOAM visual plus lidar odometry)本方法原文提出 | 0.31 | (Zhang & Singh, 2015, Table I, Sec. VII-A) |
Zhang & Singh, 2015 · Table II 本方法 8 筆
指標Relative position error
表格設定（擷取紀錄原文）：Relative position errors in fast motion tests; 'Failed' = visual features lost tracking during fast turns (Zhang & Singh, 2015, Table II)
Relative position error，authors' custom camera-lidar sensor data · Test 4, 66 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。
- 失敗
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 & Singh, 2015 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 & Singh, 2015, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| V-LOAM W-S (wide-angle camera, slow motion)本方法原文提出 | 0.67 | (Zhang & Singh, 2015, Table II, Sec. VII-B) |
| V-LOAM Fi-S (fisheye camera, slow motion)本方法原文提出 | 0.68 | (Zhang & Singh, 2015, Table II, Sec. VII-B) |
| V-LOAM W-Fa (wide-angle camera, fast motion)本方法原文提出 | 無數值失敗註記（擷取紀錄）：failed | (Zhang & Singh, 2015, Table II, Sec. VII-B) |
| V-LOAM Fi-Fa (fisheye camera, fast motion)本方法原文提出 | 1.3 | (Zhang & Singh, 2015, Table II, Sec. VII-B) |
Yuan et al., 2023a · Table VI (KITTI part) 本方法 1 筆
指標Relative translational error (RTE)
資料集與序列KITTI odometry · 11-21 mean (KITTI test set)
表格設定（擷取紀錄原文）：KITTI odometry; relative translational error (%) of visual-LiDAR odometry; '+' = open-source LiDAR odometry modified by the authors to use the SDV-LOAM visual module as motion prior; V-LOAM only has test-set results; '-' cells not stored (Yuan et al., 2023a, Table VI (KITTI part))
Relative translational error (RTE)，KITTI odometry · 11-21 mean (KITTI test set)
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 VI (KITTI part))
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| V-LOAM本方法 | 0.54% | (Yuan et al., 2023a, Table VI; Sec. VII) |
| Ours (SDV-LOAM)原文提出 | 0.6% | (Yuan et al., 2023a, Table VI; Sec. VII) |
Zhang & Singh, 2015 · Text Sec. VIII 本方法 1 筆
指標relative position drift (benchmark ranking by average translation and rotation errors)
資料集與序列KITTI odometry benchmark · benchmark average
表格設定（擷取紀錄原文）：KITTI odometry benchmark result as stated by the authors (ranked first at the time); no per-sequence table in the paper (Zhang & Singh, 2015, Text Sec. VIII)
relative position drift (benchmark ranking by average translation and rotation errors)，KITTI odometry benchmark · benchmark average
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 & Singh, 2015 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 |
|---|---|---|
| V-LOAM本方法原文提出 | 0.75% | (Zhang & Singh, 2015, Abstract, Sec. VIII) |
其他比較組
來源
Zhang & Singh, 2015
（2015）Visual-lidar odometry and mapping: low-drift, robust, and fast2015 IEEE International Conference on Robotics and Automation (ICRA), pp. 2174-2181
同儕審查已出版已讀全文經典查證後修正