SLAM 點雲圖鑑Point clouds for construction
M-01 · 2020

SegMap

SegMap 把 LiDAR 點雲切成可重複擷取的片段（segment），每個片段以 CNN 壓縮成 64 維描述子，再以描述子的最近鄰檢索加上片段質心的幾何一致性檢查，得到相對於地圖的六自由度定位。這些定位結果作為迴圈約束，與 ICP 或 LOAM 里程計一起放入 iSAM2 位姿圖，也能在多機器人之間合併地圖。描述子訓練同時兼顧檢索與重建，因此可由描述子重建近似點雲或網格，並分辨車輛、建物與其他類別，以剔除可能移動的物體。

具全域修正的完整 SLAM近十年已讀全文同儕審查已出版LOAM 系列與特徵式 LiDAR 建圖

閱讀版本
arXiv v1 (2019-09-27), accepted IJRR manuscript (sagej template); SAGE version of record not read
全文取得途徑
arXiv 預印本
本頁內容

Segment-based LiDAR mapping and global localization: repeatable point-cloud segments are encoded by a 64-D learned descriptor, matched by k-NN retrieval plus centroid geometric verification, and fed as loop-closure or multi-robot constraints into an iSAM2 pose graph with ICP or LOAM odometry; the same descriptor supports map reconstruction and semantic filtering.

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 (KITTI odometry; sensor model not named in the paper)、rotating 2D SICK LMS-151 LiDAR on UGVs that also carried motor encoders and an Xsens MTI-G IMU (search and rescue experiments)
原文測試平台vehicle (KITTI)、UGV (search and rescue; locomotion type not stated)
狀態估計incremental pose-graph SLAM (iSAM2) that combines LiDAR odometry (ICP-based, or LOAM loosely coupled) with 6-DoF segment-based localization constraints; multi-robot variant runs one centralized pose graph (Sec. 3; Sec. 5.1; Sec. 5.8; Sec. 5.9)
資料關聯segments grown incrementally in a dynamic voxel grid (Euclidean clustering after ground removal, or smoothness-based planar growing); 64-D CNN descriptor per segment; k-NN retrieval in descriptor space (64 neighbours in the LOAM experiment) and geometric consistency of segment centroids with at least 7 correspondences (Sec. 3; Sec. 4.1; Sec. 5.8)
時間表示discrete poses (pose graph)
去畸變in the LOAM plus SegMap system LOAM undistorts the scans (Sec. 5.8); otherwise not described
迴圈閉合segment descriptor retrieval with centroid geometric verification, used for loop closures and for multi-robot global associations; semantic filtering can reject vehicle segments (Sec. 3; Sec. 5.8; Sec. 5.9.1)
全域最佳化incremental pose-graph optimization based on iSAM2 (Sec. 5.1)
地圖表示target map of segment centroids with 64-D descriptors (only the last and most complete observation kept); local map radius 50 m; a decoder reconstructs voxel point clouds or marching-cubes meshes from the descriptors (Sec. 3; Sec. 4.3; Sec. 5.5; Sec. 5.8)
先驗資訊optional prior segment map for global localization; descriptor network trained on KITTI sequences 05 and 06 and reused unchanged indoors (Sec. 5.3; Sec. 5.9)
可輸出幾何robot trajectories, a compact segment map, and point clouds or meshes reconstructed from descriptors (Figs. 5, 9, 12, 14)
計算需求Intel i7-6700K with NVIDIA GeForce GTX 980 Ti; one descriptor takes 0.8 ms on GPU (SegMini 0.3 ms) and 245 ms on CPU (SegMini 41 ms); in the five-robot KITTI run localization and reconstruction ran at 10.5 Hz on average with 28.4 ms per local cloud for description (Sec. 5.1; Sec. 5.3; Sec. 5.9.1)

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
LiDARSICK LMS-151 (rotating 2D)方法輸入search and rescue UGV data (Gustav Knepper powerplant; Phoenix-West foundry)rotating 2D LiDAR on UGVs; field of view not full 360 degrees(Dubé et al., 2020, Sec. 5.9.2)
慣性量測單元（IMU）Xsens MTI-G方法輸入search and rescue UGV data (Gustav Knepper powerplant; Phoenix-West foundry)原文未報告(Dubé et al., 2020, Sec. 5.9.2)
GNSS 接收器GPS (KITTI; model not named)參考或真值量測KITTI odometryused to find ground-truth segment correspondences in revisited areas(Dubé et al., 2020, Sec. 5.2.2)
輪式或腿式里程計UGV motor encoders (model not reported)方法輸入search and rescue UGV data (Gustav Knepper powerplant; Phoenix-West foundry)multiple motor encoders(Dubé et al., 2020, Sec. 5.9.2)
載具平台UGV (model not reported)方法輸入search and rescue UGV data (Gustav Knepper powerplant; Phoenix-West foundry)three UGVs in the powerplant, two in the foundry(Dubé et al., 2020, Sec. 5.9.2; Table 2)
運算硬體Intel i7-6700K執行運算平台未標示processor used for all experiments(Dubé et al., 2020, Sec. 5.1)
運算硬體NVIDIA GeForce GTX 980 Ti執行運算平台未標示GPU used for all experiments (TensorFlow)(Dubé et al., 2020, Sec. 5.1)

作者報告的優勢與限制

優勢

  • Learned descriptor retrieves correct segments with far fewer neighbours than an eigenvalue-based handcrafted baseline over most of the segment-growing process (Fig. 8; Sec. 5.4)
  • Localizes about 6% more KITTI 00 queries than LocNet and returns full 6-DoF poses (Sec. 5.7; Fig. 10)
  • Combined with LOAM, segment-based loop closures reduce long-path drift on KITTI 00 and 08 by up to about half (Sec. 5.8; abstract)
  • Descriptor maps are compact: the KITTI 00 map shrinks from 16.8 MB of raw segments to 386.2 kB (Sec. 5.9.1; Table 2)
  • The KITTI-trained model still produced localizations indoors with a non-360-degree rotating 2D LiDAR in a powerplant and a foundry, including opposite-direction revisits (Sec. 5.9.2)

限制

  • Relies only on geometry, so repetitive man-made structures can cause perceptual aliasing (Sec. 6)
  • Featureless places such as flat fields or straight corridors do not allow reliable segment extraction, and the map drifts until a distinct area is reached (Sec. 6)
  • Euclidean segmentation suits outdoor scenes and curvature-based segmentation suits indoor scenes; they are not run jointly (Sec. 6)
  • Descriptor-level localization is coarse; ICP on raw clouds would refine it by 0.13 +/- 0.06 m but requires keeping and transmitting raw data (Sec. 5.9.1)
  • CPU descriptor extraction is slow (245 ms per descriptor) without a GPU (Sec. 5.3)

營建工程相關證據

論文未在施工現場測試；搜救實驗是以 UGV 在停用發電廠的兩層建築與半開放鋼構鑄造廠內建圖，屬既有建築與工業設施。一篇基礎設施非破壞檢測回顧的作者表示無法執行其程式碼（Ghadimzadeh Alamdari et al., 2025）。片段描述子可大幅壓縮地圖並支援多機器人合併，對大型工地多機協作建圖可能有參考價值，但這屬推論。

原文驗證環境：公開基準、已完工建築、基礎設施

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 24 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。

Dubé et al., 2020 · Table 2 本方法 9 筆

表格設定（擷取紀錄原文）：Statistics of the three multi-robot experiments: KITTI 00 (5 robots, 114 s), Gustav Knepper powerplant (3 UGVs, 850 s), Phoenix-West foundry (2 UGVs, 1086 s) (Dubé et al., 2020, Table 2)

Bandwidth for transmitting descriptors (kB/s)，KITTI odometry · KITTI

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 Dubé et al., 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。

統計量：原文未報告；對齊方式：不適用；單位：kB/s；場景：vehicle, urban (simulated multi-robot)

數值與出處
方法（原文寫法）報告值出處
SegMap descriptors本方法原文提出60.4 kB/s(Dubé et al., 2020, Table 2)

Dubé et al., 2020 · Table 1 本方法 4 筆

指標Average ratio of corresponding points within one voxel distance

資料集與序列KITTI odometry · 00 (segments)

表格設定（擷取紀錄原文）：Average ratio of corresponding points within one voxel between original and reconstructed segments (KITTI 00 segments), by descriptor size (Dubé et al., 2020, Table 1)

Average ratio of corresponding points within one voxel distance，KITTI odometry · 00 (segments)

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。

這些是 Dubé et al., 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。

統計量：平均值（mean）；對齊方式：不適用；單位：ratio；場景：vehicle, urban

資料來源作者報告值(Dubé et al., 2020, Table 1)

數值與出處
方法（原文寫法）報告值出處
Autoencoder baseline, descriptor size 160.87 ratio(Dubé et al., 2020, Table 1)
SegMap, descriptor size 16本方法原文提出0.86 ratio(Dubé et al., 2020, Table 1)
Autoencoder baseline, descriptor size 320.91 ratio(Dubé et al., 2020, Table 1)
SegMap, descriptor size 32本方法原文提出0.89 ratio(Dubé et al., 2020, Table 1)
Autoencoder baseline, descriptor size 640.93 ratio(Dubé et al., 2020, Table 1)
SegMap, descriptor size 64本方法原文提出0.91 ratio(Dubé et al., 2020, Table 1)
Autoencoder baseline, descriptor size 1280.94 ratio(Dubé et al., 2020, Table 1)
SegMap, descriptor size 128本方法原文提出0.92 ratio(Dubé et al., 2020, Table 1)

Dubé et al., 2020 · Text Sec. 5.3 本方法 4 筆

指標time to compute a descriptor

資料集與序列KITTI odometry

表格設定（擷取紀錄原文）：Average time to compute one segment descriptor (Dubé et al., 2020, Text Sec. 5.3)

time to compute a descriptor，KITTI odometry

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。

這些是 Dubé et al., 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。

統計量：平均值（mean）；對齊方式：不適用；單位：ms

資料來源作者報告值(Dubé et al., 2020, Text Sec. 5.3)

數值與出處
方法（原文寫法）報告值出處
SegMap descriptor (GPU)本方法原文提出硬體：NVIDIA GeForce GTX 980 Ti0.8 ms(Dubé et al., 2020, Sec. 5.3)
SegMini descriptor (GPU)本方法原文提出硬體：NVIDIA GeForce GTX 980 Ti0.3 ms(Dubé et al., 2020, Sec. 5.3)
SegMap descriptor (CPU)本方法原文提出硬體：Intel i7-6700K245 ms(Dubé et al., 2020, Sec. 5.3)
SegMini descriptor (CPU)本方法原文提出硬體：Intel i7-6700K41 ms(Dubé et al., 2020, Sec. 5.3)

Dubé et al., 2020 · Text Sec. 5.9.1 本方法 3 筆

資料集與序列KITTI odometry · 00 (five-robot split)

表格設定（擷取紀錄原文）：KITTI 00 split into five simultaneously played robots (114 s), vehicle segments rejected (Dubé et al., 2020, Text Sec. 5.9.1)

average frequency of localization and map reconstruction，KITTI odometry · 00 (five-robot split)

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 Dubé et al., 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。

統計量：平均值（mean）；對齊方式：不適用；單位：Hz；場景：vehicle, urban

數值與出處
方法（原文寫法）報告值出處
SegMap本方法原文提出硬體：single machine with Intel i7-6700K and NVIDIA GeForce GTX 980 Ti10.5 Hz(Dubé et al., 2020, Sec. 5.9.1)

其他比較組

列出其餘 2 個比較組

來源

  • Dubé et al., 2020

    Renaud Dubé, Andrei Cramariuc, Daniel Dugas, Hannes Sommer, Marcin Dymczyk, Juan Nieto, Roland Siegwart, Cesar Cadena（2020）SegMap: Segment-based mapping and localization using data-driven descriptorsThe International Journal of Robotics Research, 39(2-3):339-355

    DOI 10.1177/0278364919863090arXiv 1909.12837程式碼

    同儕審查已出版已讀全文近十年

程式碼：https://github.com/ethz-asl/segmap（授權：BSD-3-Clause (LICENSE file read)）。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。

回到方法圖鑑

SLAM 點雲圖鑑

本站為綜述論文《SLAM-Based Point Clouds for Construction: A Review》研究過程的互動說明文件。互動圖為教學模擬，不代表任何特定系統的實測表現；方法描述、數值與結論以所引原始文獻為準。

內容基準日 2026-09-25 · 研究與網站製作使用 AI 輔助工具，見 證據標記與方法

選擇開啟Esc關閉