STD
STD 在由數次掃描累積而成的關鍵影格上，先以體素共變異數矩陣的特徵值判斷平面並以區域成長擴展，再把平面邊界體素中的點投影到所屬平面形成影像，取 5×5 鄰域極大值作為關鍵點；每個關鍵點以 kd-tree 取 20 個近鄰組成三角形，三邊長與三個法向量內積共六個屬性對剛體變換不變，作為雜湊鍵投票檢索前 10 個候選關鍵影格；再以三角形頂點對應經 SVD 與 RANSAC 求相對位姿，並以平面重合比例做幾何驗證，可選擇以 STD-ICP 精化位姿。方法支援非重複掃描的固態光達。
本頁內容
STD extracts plane-boundary keypoints from accumulated keyframes, encodes them as rigid-invariant triangles stored in a hash table, and verifies loops geometrically, supporting non-repetitive solid-state LiDARs.
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 |
|---|---|
| 原文測試平台 | vehicle、未查證 |
| 狀態估計 | 不適用 for odometry; the loop relative pose is solved in closed form by SVD from the three matched triangle vertices inside RANSAC (maximizing correctly matched descriptors), optionally refined by STD-ICP, a Ceres optimization of plane normal difference and point-to-plane distance between coinciding planes |
| 資料關聯 | triangle descriptors from keypoints on plane boundaries; hash table on rotation/translation-invariant side lengths and normal dot products; top-10 candidates; RANSAC and plane-based geometric verification |
| 時間表示 | 不適用 |
| 去畸變 | 原文未報告 (motion compensation is not discussed); the component receives scans already registered by an external LiDAR odometry and accumulates 10 scans per keyframe for spinning LiDARs and 20 for Livox LiDARs |
| 迴圈閉合 | hash-table voting selects the top-10 candidate keyframes; a candidate is accepted when the plane coincidence percentage after the RANSAC transform exceeds sigma_pc (0.5 to 0.6 suggested from KITTI08); returns a 6-DoF relative pose, with loop correction left to the host SLAM back-end |
| 全域最佳化 | none (component) |
| 地圖表示 | per-keyframe triangle descriptor hash database plus extracted planes |
| 先驗資訊 | none; relies on an external LiDAR odometry to register scans into keyframes (text cites ref. [28], the adaptive voxel map odometry of Yuan et al. 2022; the Fig. 2 input block is labelled LiDAR odometry and mapping (LOAM)) |
| 可輸出幾何 | 不適用 |
| 計算需求 | All experiments run on one system with an Intel i7-11700k @ 3.6 GHz and 16 GB memory (no GPU is mentioned); per-frame time does not grow linearly with the number of stored frames because of the hash table, and is similar to M2DP while processing 10 times more points on KITTI00; STD and STD-ICP take less than 1% of GICP time for loop-node registration |
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 | Livox Avia | 方法輸入 | self-collected Avia Park1, Avia Park2 and Avia Indoor (multi-floor building) | small-FOV, non-repetitive scanning solid-state LiDAR; 20 frames accumulated per keyframe | (Yuan et al., 2023b, Abstract; Fig. 1; Sec. IV-B; Fig. 12) |
| LiDAR | Livox Horizon | 資料集感測器 | KA Urban East (open-sourced with LiLi-OM) | solid-state LiDAR | (Yuan et al., 2023b, Sec. IV-B) |
| LiDAR | mechanical spinning LiDARs (models not reported) | 資料集感測器 | KITTI odometry, NCLT, Complex Urban | different numbers of scanning lines; 10 frames accumulated per keyframe | (Yuan et al., 2023b, Sec. IV-A) |
| 運算硬體 | Intel i7-11700k | 執行運算平台 | 未標示 | @ 3.6 GHz with 16 GB memory; same system for all experiments | (Yuan et al., 2023b, Sec. IV) |
作者報告的優勢與限制
優勢
- Works on spinning and non-repetitive solid-state LiDARs with the same parameters except indoor voxel size (Sec. IV-A, IV-B)
- Descriptor matching gives point correspondences for geometric verification and relative pose (abstract; Sec. III)
- Gives a full 6-DoF relative pose; on KITTI00 loop nodes with perturbed initial values, STD-ICP reached accuracy similar to GICP with lower variance at less than 1% of GICP time (Sec. IV-A4, Fig. 10)
- Hash-table database keeps query time from growing linearly with database size, unlike Scan Context and M2DP (Sec. IV-A2, Fig. 8)
- Detected loops in narrow scenes with little vertical variation where Scan Context is reported to struggle (Sec. IV-A1, Fig. 7a)
限制
- Indoor multi-floor building dataset showed relatively low precision and recall because corridors on each floor are very similar (Sec. IV-B1)
- Sec. IV-A states that Scan Context and M2DP results on the public datasets were taken directly from the Scan Context paper, so conditions are not identical, while Sec. IV-A2 says the modified Scan Context MATLAB code (8 augmentations) was used to obtain the Sec. IV-A1 results; the source of the baseline curves is therefore internally inconsistent (Sec. IV-A; Sec. IV-A2)
- Ground truth for the Livox datasets was built from LiDAR-inertial odometry plus loop closure and pose graph, not an independent reference (Sec. IV-B)
- Authors state STD performs poorly only when structures or planes are particularly sparse because extracted keypoints become scarce (Sec. IV-A1, Fig. 7b, NCLT)
- Scan Context was not compared on the Livox datasets because the authors consider it incompatible with Livox solid-state LiDARs; indoor true-positive threshold was 4 m versus 20 m outdoors (Sec. IV-B)
- Runtime baselines ran from MATLAB code (M2DP default, Scan Context modified with 8 augmentations) while the STD implementation language is not stated, so timing parity is unclear (Sec. IV-A2; parity point is an inference)
- Results are reported only as plots (precision-recall, time, error), with no tabulated values (Figs. 6, 8-10, 12)
營建工程相關證據
無工地測試；作者以 Livox Avia 蒐集多樓層建築室內資料，指出各樓層走廊相似導致精確率與召回率較低，但仍能提供一定數量的有效迴圈，作者認為可用於多樓層停車場、博物館等室內建圖（Sec. IV-B1）。營建中重複樓層與標準化空間為直接相關風險（推論）；LTA-OM 以 STD 為迴圈偵測並在多層相似結構建築驗證（Zou et al., 2024 abstract）。
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 1 個比較組，合計 9 筆紀錄。
Lim et al., 2025 · Table I 本方法 9 筆
資料集與序列KITTI · 10 m benchmark
表格設定（擷取紀錄原文）：KITTI 10 m benchmark [23]: scan-to-scan global registration; success if translation < 2 m and rotation < 5 deg; RTE and RRE averaged over successful registrations only (Sec. IV-A); W = submap window size; learning-based results as listed by the authors (Lim et al., 2025, Table I)
RTE [cm]，KITTI · 10 m benchmark
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 Lim et al., 2025 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(Lim et al., 2025, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| 3DFeat-Net | 25.9 cm | (Lim et al., 2025, Table I) |
| FCGF | 6.47 cm | (Lim et al., 2025, Table I) |
| DIP | 8.69 cm | (Lim et al., 2025, Table I) |
| Predator | 5.6 cm | (Lim et al., 2025, Table I) |
| SpinNet | 9.88 cm | (Lim et al., 2025, Table I) |
| D3Feat | 11 cm | (Lim et al., 2025, Table I) |
| GeDi | 7.55 cm | (Lim et al., 2025, Table I) |
| G-ICP | 8.56 cm | (Lim et al., 2025, Table I) |
| STD, W = 1本方法 | 26.09 cm | (Lim et al., 2025, Table I) |
| STD, W = 3本方法 | 20.94 cm | (Lim et al., 2025, Table I) |
| STD, W = 5本方法 | 23.97 cm | (Lim et al., 2025, Table I) |
| MapClosures, W = 1 | 38.5 cm | (Lim et al., 2025, Table I) |
| MapClosures, W = 3 | 31.5 cm | (Lim et al., 2025, Table I) |
| MapClosures, W = 5 | 32.27 cm | (Lim et al., 2025, Table I) |
| FPFH + FGR | 6.94 cm | (Lim et al., 2025, Table I) |
| FPFH + TEASER++ | 9.36 cm | (Lim et al., 2025, Table I) |
| FPFH + Quatro | 13.15 cm | (Lim et al., 2025, Table I) |
| Proposed原文提出 | 18.1 cm | (Lim et al., 2025, Table I) |
| FPFH + FGR + G-ICP | 1.22 cm | (Lim et al., 2025, Table I) |
| FPFH + TEASER + G-ICP | 1.1 cm | (Lim et al., 2025, Table I) |
| FPFH + Quatro + G-ICP | 1.1 cm | (Lim et al., 2025, Table I) |
| Proposed + G-ICP原文提出 | 1.1 cm | (Lim et al., 2025, Table I) |
來源
Yuan et al., 2023b
（2023）STD: Stable Triangle Descriptor for 3D place recognition2023 IEEE International Conference on Robotics and Automation (ICRA), pp. 1897-1903
DOI 10.1109/icra48891.2023.10160413arXiv 2209.12435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
- 程式碼釋出：hku-mars/STD https://github.com/hku-mars/STD
程式碼：https://github.com/hku-mars/STD（授權：GPL-2.0 per README, free only for personal and academic use; commercial use requires negotiation）。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。