LOL
LOL 只用 LiDAR 在既有 3D 點雲地圖中做里程計與定位：以 LOAM 連續估計位姿，並把最近數幀點雲累積成局部地圖、切成片段，以 SegMatch 或 SegMap 描述子與預先切割描述的目標地圖片段比對。為減少誤匹配，只在里程計位置附近搜尋候選，再以質心平移一致性與 RANSAC 過濾；通過後先以匹配片段質心的平均位移做初值，再以 ICP 對齊片段點雲求出修正量，送入 SegMap 的增量位姿圖。它是在先驗地圖中定位，而不是線上迴圈閉合。
本頁內容
LiDAR-only odometry and localization in a prior point-cloud map: LOAM odometry plus SegMap/SegMatch segment matching against a pre-segmented target map, with location-gated candidate search, centroid consistency and RANSAC filtering, and a centroid-shift prior refined by ICP; updates enter SegMap's incremental pose graph.
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 | 3D LiDAR only (KITTI raw drives; sensor model not named in the paper); no IMU, wheel encoder or GPS (Sec. I) |
|---|---|
| 原文測試平台 | vehicle (KITTI raw drives 18, 27, 28) |
| 狀態估計 | LOAM odometry and mapping (mapping at one tenth of the odometry rate) supplies poses; relocalization updates from segment matches are inserted into SegMap's incremental pose-graph mapping module, modified to incorporate relocalization in the global map (Sec. III-A; Sec. III-B; Fig. 1) |
| 資料關聯 | SegMatch eigenvalue (1x7) or SegMap CNN (1x64) segment descriptors; target segments searched only within a distance threshold of the odometry position; pairwise centroid-translation consistency, RANSAC alignment of matched centroids, a mean centroid-shift prior and final ICP between matched segment point clouds (Sec. III-C; Sec. III-D) |
| 時間表示 | discrete poses |
| 去畸變 | not described (LOAM front end used as is) |
| 迴圈閉合 | not used; drift is cancelled by relocalization against the prior target map instead of loop closure in the online map (Sec. I) |
| 全域最佳化 | SegMap incremental pose-graph mapping module receives the relocalization updates (Sec. III-A; Fig. 1) |
| 地圖表示 | LOAM map voxelized at 5 cm; online local cloud densified from the last k scans and segmented; offline target point cloud segmented and described into a segment database searchable by centroid position (Sec. III-A; Sec. III-B) |
| 先驗資訊 | prior 3D point cloud target map (built with the laser SLAM tool released with SegMap in the experiments) and an assumed start position before the first localization (Sec. III-A; Sec. IV) |
| 可輸出幾何 | trajectory relocalized in the target map frame |
| 計算需求 | Intel i7-6700K, 32 GB RAM, NVIDIA GeForce GTX 1080; mean times 372.7 ms segmentation, 0.40 ms description, 26.27 ms match recognition, 0.09 ms RANSAC filtering and 88.0 ms ICP alignment; the authors describe the system as real time (Sec. IV) |
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 | 3D LiDAR of the KITTI raw recordings (model not named in the paper) | 資料集感測器 | KITTI raw drives 18, 27, 28 | raw drives 18 (about 2200 m), 27 (about 3660 m) and 28 (about 4125 m) | (Rozenberszki & Majdik, 2020, Sec. IV; Fig. 2) |
| 運算硬體 | Intel i7-6700K | 執行運算平台 | 未標示 | 32 GB RAM | (Rozenberszki & Majdik, 2020, Sec. IV) |
| 運算硬體 | Nvidia GeForce GTX 1080歸入：NVIDIA GeForce GTX1080 | 執行運算平台 | 未標示 | GPU in the same system | (Rozenberszki & Majdik, 2020, Sec. IV) |
| 其他 | ground-truth target maps built with the laser SLAM tool released with SegMap | 參考或真值量測 | KITTI raw drives 18, 27, 28 | used as target maps and trajectory reference | (Rozenberszki & Majdik, 2020, Sec. IV) |
作者報告的優勢與限制
優勢
- With a minimum cluster size of 3 or more, the filtering produced at most one false-positive match per drive while keeping 14 to 70 true positives (Table I)
- Trajectories on three long KITTI raw drives (about 2.2 to 4.1 km) are significantly better than LOAM alone and relocalization never lost tracking (Sec. III-D; Sec. IV; Figs. 2 and 4)
- No IMU, wheel encoder or GPS required (Sec. I)
- Open-source code released (Supplementary attachments)
限制
- Quantitative trajectory error is only shown as plots for Drive 18 (Fig. 4); no numeric trajectory table
- Evaluated on three KITTI raw drives in residential and city areas only (Sec. IV)
- Target maps were built with the SegMap laser SLAM tool, not an independent survey (Sec. IV)
- Cluster size 2 still gives incidental false localizations and size 5 gives too few localizations (Sec. IV)
- Needs an assumed start position before the first localization (Sec. III-A)
營建工程相關證據
在既有點雲地圖中定位的概念，與施工現場以先前掃描或 BIM 衍生點雲作為參考地圖相近，但這屬推論；論文只在 KITTI 市區與住宅區道路測試，目標地圖以 SegMap 的雷射 SLAM 工具建立，並非測量等級參考。一篇基礎設施非破壞檢測回顧的作者回報 LOL 無法完成軌跡估計（Ghadimzadeh Alamdari et al., 2025）。
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 41 筆紀錄。
Rozenberszki & Majdik, 2020 · Table I 本方法 36 筆
表格設定（擷取紀錄原文）：Numbers of filtered-out, true-positive and false-positive SegMap matches for minimum cluster sizes 2-5 on three KITTI raw drives; validity judged against the ground-truth trajectory (Rozenberszki & Majdik, 2020, Table I)
number of filtered out matches，KITTI raw · Drive 18
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 Rozenberszki & Majdik, 2020 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(Rozenberszki & Majdik, 2020, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LOL filtering, minimum cluster size 2本方法原文提出 | 579 count | (Rozenberszki & Majdik, 2020, Table I) |
| LOL filtering, minimum cluster size 3本方法原文提出 | 299 count | (Rozenberszki & Majdik, 2020, Table I) |
| LOL filtering, minimum cluster size 4本方法原文提出 | 8 count | (Rozenberszki & Majdik, 2020, Table I) |
| LOL filtering, minimum cluster size 5本方法原文提出 | 0 count | (Rozenberszki & Majdik, 2020, Table I) |
Rozenberszki & Majdik, 2020 · Text Sec. IV 本方法 5 筆
資料集與序列KITTI raw · drives 18, 27, 28
表格設定（擷取紀錄原文）：Mean (standard deviation) processing time of each module (Rozenberszki & Majdik, 2020, Text Sec. IV)
segmentation time, mean (std 7.2)，KITTI raw · drives 18, 27, 28
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 Rozenberszki & Majdik, 2020 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 |
|---|---|---|
| LOL, segmentation本方法原文提出硬體：Intel i7-6700K, 32 GB RAM, Nvidia GeForce GTX 1080 | 372.7 ms | (Rozenberszki & Majdik, 2020, Sec. IV) |
來源
Rozenberszki & Majdik, 2020
（2020）LOL: Lidar-only Odometry and Localization in 3D point cloud maps2020 IEEE International Conference on Robotics and Automation (ICRA), pp. 4379-4385
DOI 10.1109/icra40945.2020.9197450arXiv 2007.01595程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 預印本：LOL (arXiv v1, ICRA 2020 preprint version) https://arxiv.org/abs/2007.01595
- accepted manuscript：SZTAKI eprint 10046 (listed by OpenAlex; not opened) http://eprints.sztaki.hu/10046/
- 程式碼釋出：RozDavid/LOL https://github.com/RozDavid/LOL
程式碼：https://github.com/RozDavid/LOL（授權：not stated (no license file found at the repository root; README has no license statement)）。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。