Scan Context
Scan Context 以感測器為中心,將單次 3D 光達掃描劃分為 20 個環(ring)乘 60 個扇區(sector)的極座標格網(最大距離 80 m),每格記錄其中點的最大高度,形成 2D 全域描述子,不依賴直方圖或事前訓練。搜尋分兩階段:先以各環佔有率組成的環鍵(ring key)建 kd-tree 取出 10 或 50 個候選,再對候選做所有欄位平移的逐欄餘弦距離比較,最小距離低於門檻即判定為迴圈;平移量同時給出約 6° 解析度的偏航初值,可供 ICP 使用,因此反向重訪與轉角也能偵測迴圈。
本頁內容
Scan Context encodes a LiDAR scan as an egocentric polar grid descriptor and detects loops by ring-key kd-tree retrieval plus column-shift (yaw-aligned) comparison, without training.
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 (Velodyne HDL-64E on KITTI, HDL-32E on NCLT, two tilted VLP-16 merged on Complex Urban LiDAR) |
|---|---|
| 原文測試平台 | vehicle (KITTI: HDL-64E located in the centre of the car)、Segway mobile platform (NCLT)、Complex Urban LiDAR dataset (two tilted VLP-16; platform not described in the paper) |
| 狀態估計 | 不適用 |
| 資料關聯 | egocentric polar grid Nr = 20 rings x Ns = 60 sectors, Lmax = 80 m, bin value = maximum point height (Eq. 3), empty bins 0; optional root-shift augmentation with Ntrans = 8 translated copies for lane-level offsets; ring key = per-ring occupancy ratio (L0 norm) indexed in a KD tree, 10 or 50 candidates; column-wise cosine distance minimised over all column shifts, accepted below threshold tau; 0.6 m grid downsampling |
| 時間表示 | 不適用 |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | provides loop candidates with coarse yaw alignment (6 deg resolution); the yaw shift initialises point-to-point ICP, which reduced ICP time and RMSE for KITTI 08 reverse loops (Fig. 7, 8); pose-graph use left to host SLAM |
| 全域最佳化 | none (component) |
| 地圖表示 | per-keyframe 2D descriptor database |
| 先驗資訊 | none; no training required (abstract) |
| 可輸出幾何 | 不適用 |
| 計算需求 | Matlab on Intel i7-6700 3.40 GHz with 16 GB memory; on KITTI 00, 0.1291 s per descriptor with augmentation (0.0143 s without) and 0.0807 s or 0.3331 s loop search for 10 or 50 candidates; about 2 to 5 Hz overall (Table II, Sec. IV-D) |
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 | Velodyne HDL-64E | 資料集感測器 | KITTI | 64-ray, located in the centre of the car | (Kim & Kim, 2018, Sec. IV-A1) |
| LiDAR | Velodyne HDL-32E | 資料集感測器 | NCLT | 32-ray, attached to a Segway mobile platform | (Kim & Kim, 2018, Sec. IV-A2) |
| LiDAR | Velodyne VLP-16 PUCK (two, tilted left and right)歸入:Velodyne VLP-16 | 資料集感測器 | Complex Urban LiDAR | two tilted units without 360 deg surround view; clouds merged into one scan | (Kim & Kim, 2018, Sec. IV-A3) |
| 載具平台 | Segway mobile platform | 資料集感測器 | NCLT | 原文未報告 | (Kim & Kim, 2018, Sec. IV-A2) |
| 運算硬體 | Intel i7-6700 CPU | 執行運算平台 | 未標示 | 3.40 GHz, 16 GB memory; Matlab | (Kim & Kim, 2018, Sec. IV) |
作者報告的優勢與限制
優勢
- Viewpoint (reverse revisit) invariance through column-shift alignment; substantially outperformed M2DP, Z-projection and ESF on KITTI 08, which has only reverse loops (Sec. IV-B, Fig. 5d)
- No histogram and no training required (abstract)
- Authors report outperforming M2DP, Z-projection and ESF on the outdoor urban sequences, including the challenging Complex Urban 02, in precision-recall curves (Sec. IV-B, Fig. 5)
- Yaw estimate from column shift improves ICP success, time and RMSE on reverse loops (Sec. IV-C, Fig. 7, 8)
限制
- Authors report limited performance indoors where vertical height variation is small: low recall and precision on NCLT, whose trajectory contains narrow indoor spaces (Sec. IV-B)
- All methods degraded on Complex Urban 02 with narrow roads and repeated structures of similar height and shape (Sec. IV-B, Fig. 6)
- Ring key is less informative than the full descriptor, so only 10 candidates is vulnerable with many similar structures (Sec. IV-B)
- Loop search slower than all three compared global descriptors (M2DP, Z-projection, ESF), though within 2 to 5 Hz in Matlab (Table II, Sec. IV-D)
- Code is CC BY-NC-SA 4.0, i.e., non-commercial use only (README)
- Found in the authors' journal extension (vehicle driving data): a vehicle moving along a corridor-like place and a tall, large object (e.g., a bus) very close to the sensor caused localization failures (Kim et al., 2022b Sec. VIII-F, Fig. 22)
- Found in follow-up work: LT-SLAM with Scan Context failed to find loops in a repetitive indoor parking dataset (Yang et al., 2024 Sec. IV-B)
營建工程相關證據
未報告工地測試。原論文指出在含狹窄室內空間的 NCLT 路段,因垂直高度變化小而召回率與精確率偏低(Sec. IV-B),此點與施工中建築的走廊及樓層環境相關,但套用到工地屬推論。LT-mapper 以 Scan Context 作跨時段迴圈(Kim & Kim, 2022 Sec. V-A1),BIM-SLAM 亦採用並調整室內參數(Vega Torres et al., 2023 Sec. 4.3),因此它是營建相關多時段流程的實際依賴元件。
原文驗證環境:公開基準、跨場域
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 11 筆紀錄。
Wang et al., 2020 · Table II 本方法 6 筆
表格設定(擷取紀錄原文):KITTI sequences 00, 02 (forward and reverse revisits), 05; loop-closure precision and recall (%); Scan Context, GLAROT3D and Cieslewski results copied from their papers, DBoW2 run by the authors; loop ground truth from GPS (Wang et al., 2020, Table II)
Precision (%),KITTI odometry · sequence 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 Wang 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。
資料來源作者報告值(Wang et al., 2020, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Kim [21] (Scan Context)本方法 | 100% | (Wang et al., 2020, Table II) |
| GLAROT3D [17] | 86% | (Wang et al., 2020, Table II) |
| Cieslewski [24] | 92% | (Wang et al., 2020, Table II) |
| Galvez-Lopez [10] (DBoW2) | 100% | (Wang et al., 2020, Table II) |
| Proposed (ISC)原文提出 | 100% | (Wang et al., 2020, Table II) |
Kim & Kim, 2018 · Table II 本方法 4 筆
資料集與序列KITTI · 00
表格設定(擷取紀錄原文):Average time on KITTI 00; 0.6 m3 grid downsampling for all methods (Sec. IV-D); scan context creation includes optional root-shift augmentation; implementations: Scan Context in Matlab, M2DP authors' Matlab code, Z-projection in Matlab, ESF from PCL in C++ (Kim & Kim, 2018, Table II)
Calculating Descriptor (s),KITTI · 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 Kim & Kim, 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(Kim & Kim, 2018, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Scan context-10本方法原文提出硬體:Intel i7-6700 CPU 3.40 GHz, 16 GB memory (Matlab) | 0.1291 s | (Kim & Kim, 2018, Table II) |
| Scan context-50本方法原文提出硬體:Intel i7-6700 CPU 3.40 GHz, 16 GB memory (Matlab) | 0.1291 s | (Kim & Kim, 2018, Table II) |
| M2DP硬體:Intel i7-6700 CPU 3.40 GHz, 16 GB memory (Matlab) | 0.0218 s | (Kim & Kim, 2018, Table II) |
| Z-projection硬體:Intel i7-6700 CPU 3.40 GHz, 16 GB memory (Matlab) | 0.0472 s | (Kim & Kim, 2018, Table II) |
| ESF硬體:Intel i7-6700 CPU 3.40 GHz, 16 GB memory (PCL C++) | 0.0635 s | (Kim & Kim, 2018, Table II) |
Kim & Kim, 2018 · Text Sec.IV-D 本方法 1 筆
指標time to create a single scan context (without augmentation)
資料集與序列KITTI · 00
表格設定(擷取紀錄原文):Single scan context creation without augmentation (Kim & Kim, 2018, Text Sec.IV-D)
time to create a single scan context (without augmentation),KITTI · 00
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 Kim & Kim, 2018 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 |
|---|---|---|
| Scan context本方法原文提出硬體:Intel i7-6700 CPU 3.40 GHz, 16 GB memory (Matlab) | 0.0143 s | (Kim & Kim, 2018, Sec. IV-D) |
來源
Kim & Kim, 2018
(2018)Scan Context: Egocentric Spatial Descriptor for Place Recognition Within 3D Point Cloud Map2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 4802-4809
DOI 10.1109/iros.2018.8593953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
- 期刊延伸版:Scan Context++: Structural Place Recognition Robust to Rotation and Lateral Variations in Urban Environments (T-RO 2022) 10.1109/TRO.2021.3116424
- 程式碼釋出:gisbi-kim/scancontext https://github.com/gisbi-kim/scancontext
程式碼:https://github.com/gisbi-kim/scancontext(授權:CC BY-NC-SA 4.0 (stated in README; non-commercial))。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。