LIMO
LIMO 以單眼相機的特徵追蹤為主，LiDAR 只負責替影像特徵提供深度：先把單次掃描的 LiDAR 點投影到影像，在特徵周圍以深度直方圖切出前景點，再以面積最大的三點平面與視線求交得到特徵深度；地面上的特徵另以 RANSAC 擬合的地面平面處理，超過 30 m 的深度則捨棄。逐影格運動以 PnP 與對極誤差估計後，交給關鍵影格光束法平差，把 LiDAR 深度當成殘差加入，並以關鍵影格與地標篩選、語意剔除動態物件、植生降權及修剪最大殘差維持即時性。系統只做視覺里程計，不含迴圈閉合，也不建立稠密點雲地圖。
本頁內容
Monocular visual odometry that obtains per-feature depth from single LiDAR scans by fitting local planes to foreground-segmented projected points, then refines motion with robustified keyframe bundle adjustment that uses the LiDAR depths as residuals; evaluated on KITTI without loop closure.
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 (KITTI grayscale images used for feature tracking; camera model not named in the paper)、3D LiDAR (KITTI; model not named in the paper), used only to give depth to image features |
|---|---|
| 原文測試平台 | vehicle (KITTI odometry benchmark) |
| 狀態估計 | two separate optimizations: (i) frame-to-frame 6-DoF motion from a perspective-n-point cost plus an epipolar cost, each wrapped in a Cauchy loss, used as prior; (ii) windowed keyframe bundle adjustment over reprojection errors, LiDAR depth residuals and a scale regularizer on the oldest motion in the window, with Cauchy losses and a trimmed-least-squares-like removal of the highest residuals after a few iterations (Secs. IV-V, Eq. 7, Algorithm 1) |
| 資料關聯 | viso2 feature tracking (about 2000 correspondences in 30-40 ms); feature depth from a local plane through the maximum-area triangle of foreground LiDAR points selected in an image-space neighbourhood by a depth histogram (bin width 0.3 m); ground-plane features use a RANSAC ground fit instead; depth estimates beyond 30 m or at grazing angles rejected; landmarks on dynamic semantic classes rejected and vegetation landmarks weighted (Secs. II, III, V-C, VI) |
| 時間表示 | discrete frames and keyframes; one-shot depth from a single LiDAR scan, without accumulating point clouds over time (Sec. III) |
| 去畸變 | 原文未報告 (no LiDAR motion compensation step is described; KITTI scans are used as provided) |
| 迴圈閉合 | none; the authors state they aim for visual odometry and perform no loop closure (Fig. 2 caption; Sec. I) |
| 全域最佳化 | none |
| 地圖表示 | sparse triangulated landmarks inside the bundle-adjustment window, split into near, middle and far bins and thinned by a voxel filter with median filtering; no dense LiDAR map (Sec. V-C) |
| 先驗資訊 | LiDAR-camera calibration supplied with KITTI; semantic segmentation network (modified ResNet38) trained on Cityscapes (Sec. VI) |
| 可輸出幾何 | camera poses and a sparse landmark reconstruction; no registered point-cloud map is produced (Fig. 2 caption) |
| 計算需求 | 3.5 GHz CPU: Liviodo (frame-to-frame part) at 10 Hz on 2 cores, LIMO at 5 Hz on 4 cores; semantic labels from a modified ResNet38 at 100 ms per image on an Nvidia TitanX Pascal (Sec. VI) |
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 | KITTI LIDAR (model not named in the paper) | 資料集感測器 | KITTI odometry | LIDAR point clouds with calibration provided by KITTI; used only for feature depth | (Graeter et al., 2018, Sec. VI) |
| 相機 | KITTI camera (model not named in the paper) | 資料集感測器 | KITTI odometry | grayscale and color images provided by KITTI; grayscale used for tracking | (Graeter et al., 2018, Sec. VI) |
| 運算硬體 | Nvidia TitanX Pascal歸入：Nvidia Titan-X Pascal | 執行運算平台 | 未標示 | GPU running the modified ResNet38 semantic segmentation at 100 ms per image | (Graeter et al., 2018, Sec. VI) |
| 運算硬體 | 3.5 GHz CPU (model not reported) | 執行運算平台 | 未標示 | Liviodo on 2 cores at 10 Hz, LIMO on 4 cores at 5 Hz | (Graeter et al., 2018, Sec. VI) |
作者報告的優勢與限制
優勢
- Ranked 13th in translation error and 11th in rotation error on the KITTI odometry benchmark as of 1 March 2018, ahead of ORB-SLAM2 and Stereo LSD-SLAM (Sec. I; Sec. VII)
- KITTI evaluation set: LIMO 0.93% translation error and 0.0026 deg/m rotation error versus 1.22% and 0.0042 deg/m for the frame-to-frame part Liviodo; bundle adjustment cuts rotation error by nearly 40% (Sec. VI)
- Described by the authors as the second best LiDAR-camera method published on KITTI and the best one that does not use ICP-based LiDAR SLAM refinement (Sec. VI)
限制
- Highway scenes such as KITTI 01, 12 and 21 give few valid depth estimates because only the road is within usable range, and large optical flow and motion blur hamper tracking (Sec. VI)
- Average errors are highest at high speed and at low speed; the low-speed error is attributed to conservative standstill detection (Fig. 7 caption)
- The local plane assumption can be violated, giving imprecise depths that the backend must compensate (Fig. 8 caption)
- The optimum vegetation weight varies between sequences, which the authors read as a more complex dependence on scene content (Sec. VI)
- No loop closure and no dense map output (Sec. I; Fig. 2 caption)
營建工程相關證據
論文只在 KITTI 車載道路資料上評估，未涉及施工現場、建築構件或點雲幾何精度；輸出為稀疏地標而非點雲地圖，無法直接作為營建點雲。其參考價值在於說明以 LiDAR 為影像特徵補深度的做法；SDV-LOAM 在同一 KITTI 設定下把 LIMO 列為比較對象（Yuan et al., 2023a）。
原文驗證環境：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 2 個比較組，合計 15 筆紀錄。
Graeter et al., 2018 · Text Sec.VI 本方法 10 筆
資料集與序列KITTI odometry · evaluation (test) set
表格設定（擷取紀錄原文）：KITTI odometry benchmark evaluation set results as published on the official server (as of 1 March 2018); official KITTI metric; Liviodo is the frame-to-frame part, LIMO the full pipeline with keyframe bundle adjustment (Graeter et al., 2018, Text Sec.VI)
mean translation error (official KITTI metric)，KITTI odometry · evaluation (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 Graeter et al., 2018 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(Graeter et al., 2018, Text Sec.VI)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LIMO本方法原文提出 | 0.93% | (Graeter et al., 2018, Sec. VI) |
| Liviodo (frame-to-frame motion only)本方法原文提出 | 1.22% | (Graeter et al., 2018, Sec. VI) |
Yuan et al., 2023a · Table V 本方法 5 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 | 1.05% | (Yuan et al., 2023a, Table V; Sec. VII) |
| LIMO*本方法 | 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) |
來源
Graeter et al., 2018
（2018）LIMO: Lidar-Monocular Visual Odometry2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 7872-7879
DOI 10.1109/iros.2018.8594394arXiv 1807.07524程式碼
同儕審查已出版已讀全文近十年
相關版本
- 預印本：LIMO: Lidar-Monocular Visual Odometry (arXiv v1, accepted version) https://arxiv.org/abs/1807.07524
- 程式碼釋出：johannes-graeter/limo https://github.com/johannes-graeter/limo
程式碼：https://github.com/johannes-graeter/limo（授權：GPL-3.0 (GitHub license metadata)）。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。