SSL_SLAM
SSL_SLAM 是針對小視野、高頻率固態 LiDAR(Intel L515)設計的輕量 LiDAR 建圖定位。它先把點雲依垂直與水平角度分格並取格內平均,再以鄰域平滑度擷取邊緣與平面特徵,使特徵在大幅旋轉下仍較一致;位姿以掃描對滑動視窗局部地圖的點到邊、點到面殘差,在李群上以高斯牛頓法求解。全域地圖只用關鍵影格更新成八元樹佔據機率地圖。系統沒有迴圈閉合,在嵌入式小電腦上可即時執行。
本頁內容
Lightweight LiDAR-only odometry and mapping for small-FoV solid-state LiDAR (Intel L515): angular-grid binning with smoothness-based edge and plane features, Gauss-Newton scan-to-local-map optimization on the Lie group, and a keyframe-updated octree occupancy map; no loop closure, real time on an embedded mini PC.
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 | solid-state LiDAR only (Intel Realsense L515, 70 x 55 deg FoV, 30 Hz) |
|---|---|
| 原文測試平台 | wheeled UGV (warehouse AGV)、handheld、ground robot in a VICON room (type not stated) |
| 狀態估計 | Gauss-Newton minimization of point-to-edge and point-to-plane residuals against a sliding-window local map, with left-perturbation updates on the Lie group and a constant-velocity initial guess (Sec. III-B; Algorithm 1) |
| 資料關聯 | points binned into an M x N grid of vertical and horizontal angle cells (cell means); edge and planar features from a local smoothness over a neighbourhood lambda; 2 nearest edge points or 3 nearest planar points found in k-d trees of the local edge and planar maps of the last q frames (Sec. III-A; Sec. III-B) |
| 時間表示 | discrete poses |
| 去畸變 | not described |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | sliding-window local edge and planar feature maps for odometry; global octree with per-cell occupancy probability updated from key frames selected by translation, rotation or elapsed-time thresholds (Sec. III-B; Sec. III-C) |
| 先驗資訊 | none |
| 可輸出幾何 | trajectory and dense 3D probabilistic octree map (Figs. 1, 4 and 5) |
| 計算需求 | 31 ms per frame on a desktop with an Intel 6-core i7-8700; 42 ms per frame on an Intel NUC with an i5-10210U on the AGV; C++ with ROS Melodic on Ubuntu 18.04 (Sec. IV-A to IV-C) |
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 | Intel Realsense L515歸入:Intel RealSense L515 | 方法輸入 | 未標示 | solid-state, 30 Hz, 70 x 55 deg FoV, 0.07 deg horizontal and vertical resolution, 0.25-9 m range, 1.4 cm accuracy, 61 x 26 mm, 95 g | (Wang et al., 2021b, Table I; Sec. IV-A) |
| 行動掃描設備 | hand-held scanner (perception module, rotation platform, computing module) | 方法輸入 | 未標示 | less than 500 g; walked at normal speed indoors | (Wang et al., 2021b, Sec. IV-D; Fig. 5) |
| 載具平台 | industrial AGV with a robot arm for gripping and transporting materials | 方法輸入 | 未標示 | solid-state LiDAR mounted at the front; maximum speed 0.8 m/s | (Wang et al., 2021b, Sec. IV-C; Fig. 4) |
| 運算硬體 | Intel i7-8700 | 執行運算平台 | 未標示 | 6-core desktop CPU (VICON experiment) | (Wang et al., 2021b, Sec. IV-A) |
| 運算硬體 | Intel NUC with Intel i5-10210U | 執行運算平台 | 未標示 | mini computer on the AGV and hand-held device | (Wang et al., 2021b, Sec. IV-A) |
| 其他 | VICON system (motion capture) | 參考或真值量測 | 未標示 | ground truth in a 4 m x 4 m room | (Wang et al., 2021b, Sec. IV-A; Sec. IV-B) |
作者報告的優勢與限制
優勢
- Translational error of 5 cm in a 4 m x 4 m VICON room at 31 ms per frame, where LOAM lost tracking under fast rotation (Sec. IV-B; Fig. 3)
- Real time on an Intel NUC (42 ms per frame) on a warehouse AGV moving up to 0.8 m/s among dynamic operators and robots (Sec. IV-C)
- Mapped machine dimensions were within a few centimetres of actual sizes for two machines (Sec. IV-C)
- In the rotation test (up to 1.57 rad/s) it succeeded in 6 of 6 trials versus 1 of 6 for A-LOAM (Table II)
限制
- No loop closure or global optimization (Sec. III) (inference from the described modules)
- Quantitative accuracy only from one small VICON room trial and two machine-dimension checks (Sec. IV-B; Sec. IV-C)
- The L515 detection range is 0.25 to 9 m, which limits use in large spaces (Table I) (inference)
- Hand-held devices suffer vibration and large viewing-angle change that can cause tracking loss (Sec. IV-D)
營建工程相關證據
論文在倉庫 AGV 與室內手持掃描器上測試,並以兩台作業機台的量測尺寸與實際尺寸比對(相差不超過 3 cm),是少數以實物尺寸檢查地圖的例子,但樣本只有兩台設備。L515 的量測距離只有 0.25 至 9 m,限制了在大型工地的適用範圍;它在 VoxelMap 論文中被列為比較基準(Yuan et al., 2022)。
原文驗證環境:受控實驗、獨立參考量測、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 3 個比較組,合計 9 筆紀錄。
Wang et al., 2021b · Text Sec. IV-C 本方法 5 筆
資料集與序列own warehouse AGV data · warehouse run
表格設定(擷取紀錄原文):Warehouse AGV at up to 0.8 m/s; mapped machine size = average Euclidean distance between picked edge points; actual sizes: machine (b) 1.15 m x 1.85 m, machine (d) 1.16 m x 1.95 m (Wang et al., 2021b, Text Sec. IV-C)
average computing time,own warehouse AGV data · warehouse run
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 Wang et al., 2021b 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 |
|---|---|---|
| proposed method (SSL_SLAM)本方法原文提出硬體:Intel NUC with i5-10210U | 42 ms | (Wang et al., 2021b, Sec. IV-C) |
Wang et al., 2021b · Table II 本方法 2 筆
資料集與序列own rotation test · rotation trials
表格設定(擷取紀錄原文):Rotation test: L515 rotated randomly from horizontal at up to 1.57 rad/s and returned; tracking loss when final angle deviation exceeds 10 deg (Wang et al., 2021b, Table II)
Success (count),own rotation test · rotation trials
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., 2021b 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., 2021b, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| The proposed method本方法原文提出 | 6 count | (Wang et al., 2021b, Table II) |
| A-LOAM | 1 count | (Wang et al., 2021b, Table II) |
Wang et al., 2021b · Text Sec. IV-B 本方法 2 筆
資料集與序列own VICON room data · VICON room trial
表格設定(擷取紀錄原文):Manually driven robot in a 4 m x 4 m VICON room; LOAM configured with the L515 angles and unchanged feature numbers (Wang et al., 2021b, Text Sec. IV-B)
translational error (statistic not stated),own VICON room data · VICON room trial
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 Wang et al., 2021b 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 |
|---|---|---|
| proposed method (SSL_SLAM)本方法原文提出 | 5 cm | (Wang et al., 2021b, Sec. IV-B) |
來源
Wang et al., 2021b
(2021)Lightweight 3-D Localization and Mapping for Solid-State LiDARIEEE Robotics and Automation Letters, 6(2):1801-1807
DOI 10.1109/lra.2021.3060392arXiv 2102.03800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:Lightweight 3-D Localization and Mapping for Solid-State LiDAR (arXiv v2, RA-L accepted preprint) https://arxiv.org/abs/2102.03800
- accepted manuscript:NTU repository handle 10356/223176 (listed by OpenAlex; not opened) https://hdl.handle.net/10356/223176
- 程式碼釋出:wh200720041/ssl_slam https://github.com/wh200720041/ssl_slam
程式碼:https://github.com/wh200720041/ssl_slam(授權:GPL-3.0 (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。