Peopleremover
本法以已配準的多站或多切片點雲建立全域體素網格,每個體素只記錄有哪些掃描在其中量到點。從每個感測器原點沿視線走訪到各量測點,若某體素被其他掃描看穿為空,體素內的點即判定為動態並移除;為避免斜掃表面與取樣不均造成誤判,以最近點的「陰影」與法向量限制走訪距離,再以叢集過濾孤立誤判,並可依掃描編號做次體素移除。演算法對掃描數與點數呈線性複雜度,只有體素大小一個參數。
本頁內容
Ray traversal through a voxel occupancy grid over registered scans to detect occupancy differences and remove dynamic objects.
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 laser scans only: Riegl VZ-400 TLS for the authors' lecturehall, campus and wrzburg datasets、third-party datasets of Underwood et al. (sim, lab, carpark; sensor not stated in this paper)、mobile-mapping scan slices from an automotive production line (authors' earlier work)、stated as compatible in principle with RADAR, RGB-D or stereo point clouds (not tested) |
|---|---|
| 原文測試平台 | static terrestrial laser scanner (tripod; one test with the scanner placed on the ground)、mobile mapping scan slices (automotive production line dataset)、simulation (sim dataset) |
| 狀態估計 | 不適用 (post-registration cleaning) |
| 資料關聯 | Amanatides-Woo voxel traversal (made stricter and free of floating-point accumulation) from each sensor origin to each point; traversal aborts at voxels holding the same scan identifier; search distances clipped by point 'shadows' using normals and per-scan sphere quadtrees; small clusters of free voxels reset to static; optional sub-voxel removal by scan identifier |
| 時間表示 | 不適用 |
| 去畸變 | 不適用 |
| 迴圈閉合 | 不適用 |
| 全域最佳化 | none |
| 地圖表示 | global regular voxel grid storing only the set of scan identifiers per voxel (no point coordinates); single parameter voxel size (0.1 to 0.6 m in Table I); C++ standard library containers |
| 先驗資訊 | registered scans with known sensor origins (wrzburg registered with slam6D from 3DTK); each volume to be cleaned must be observed by at least two scans taken far enough apart in time |
| 可輸出幾何 | cleaned point cloud without dynamic objects |
| 計算需求 | offline; linear in the number of scans and points; single-threaded benchmarks although traversal parallelizes; e.g. 687 s for lecturehall (44.6 million points, 2 scans) and 13.1 h for campus (2.2 billion points, 146 scans); hardware not reported; whole point cloud held in memory |
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 |
|---|---|---|---|---|---|
| 地面雷射掃描儀(TLS) | Riegl VZ-400 | 資料集感測器 | lecturehall, campus, wrzburg | used by the authors to record the lecturehall, campus (146 scans, about 15 million points per scan) and wrzburg (6 scans, 86 million points) datasets | (Schauer & Nuchter, 2018, Sec. IX, Table I) |
作者報告的優勢與限制
優勢
- Works for mobile-mapping scan slices and terrestrial scans (abstract)
- F1 equal to Underwood et al. on sim (0.98) and lecturehall (0.96) and higher on carpark (0.83 vs 0.78) (Table I)
- Campus dataset processed in 13.1 h versus 12.8 days for pairwise comparison with Underwood et al. (Table I)
- Only one parameter (voxel size); a 17.5 cm voxel on lecturehall cut time by 18% to 567 s with F1 0.95 (Sec. IX)
- Conservative: volumes seen by a single scan are left untouched, so occlusion is not treated as change (Sec. II, III)
限制
- Runtime per iteration reported as far slower than ERASOR on SemanticKITTI (Lim et al. 2021, Table III) (cross-study)
- Lower F1 than Underwood et al. on the noisy lab dataset with very small dynamic objects (0.42 vs 0.71) (Table I, Sec. IX)
- False positives from registration errors, sensor noise, wrong normals, mirrors and transparent objects, and from the sub-voxel option (Sec. IX)
- False negatives where a volume is seen by one scan only, is shadowed by near points (scanner on the ground), the voxel is too small or ranges are far, or objects are smaller than the minimum cluster size (Sec. IX)
- Best voxel size differs per dataset and currently needs labelled training data (Sec. IX)
營建工程相關證據
未在工地測試。引言把開挖工地(excavation site)與工廠產線列為不宜為掃描而停工的場景,也提到礦場進度監測;作者先前也把方法用於汽車產線的移動測繪。對需要在施工持續進行時掃描的工地,本法可作為配準後移除人員與機具的後處理步驟(推論)。
原文驗證環境:模擬、公開基準、已完工建築、獨立參考量測
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 4 個比較組,合計 30 筆紀錄。
Lim et al., 2021 · Table II 本方法 15 筆
表格設定(擷取紀錄原文):Static-map benchmark on five manually selected SemanticKITTI frame ranges with SuMa poses; PR and RR computed voxel-wise with 0.2 voxel size for all methods; OctoMap run at 0.05 and 0.2 voxel sizes; Removert RM3 = three removal stages, RM3+RV1 adds one revert stage (Lim et al., 2021, Table II)
Preservation Rate (PR),SemanticKITTI · 00 (frames 4390-4530)
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., 2021 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., 2021, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| OctoMap - 0.05 | 76.731% | (Lim et al., 2021, Table II) |
| OctoMap - 0.2 | 34.568% | (Lim et al., 2021, Table II) |
| Peopleremover本方法 | 37.523% | (Lim et al., 2021, Table II) |
| Removert - RM3 | 85.502% | (Lim et al., 2021, Table II) |
| Removert - RM3+RV1 | 86.829% | (Lim et al., 2021, Table II) |
| ERASOR (Ours)原文提出 | 93.98% | (Lim et al., 2021, Table II) |
Schauer & Nuchter, 2018 · Table I 本方法 12 筆
表格設定(擷取紀錄原文):Best-parameter F1 and full-pipeline single-threaded runtime; Underwood et al. run on all scan pairs; no clustering or sub-voxel step; sim: 387,838 points, 8 scans, 28 pairs, Ta 1.4, Tr 0.1 m, voxel 0.6 m (Schauer & Nuchter, 2018, Table I)
F1-score (dynamic point classification),sim (Underwood et al.) · 8 scans
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 Schauer & Nuchter, 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(Schauer & Nuchter, 2018, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| underwood | 0.98 fraction | (Schauer & Nuchter, 2018, Table I) |
| 3dtk (peopleremover)本方法原文提出 | 0.98 fraction | (Schauer & Nuchter, 2018, Table I) |
Schauer & Nuchter, 2018 · Text Sec. IX 本方法 2 筆
資料集與序列lecturehall (own, Riegl VZ-400) · 2 scans
表格設定(擷取紀錄原文):Quality-runtime trade-off on lecturehall with a 17.5 cm voxel instead of 10 cm (Schauer & Nuchter, 2018, Text Sec. IX)
F1-score (dynamic point classification),lecturehall (own, Riegl VZ-400) · 2 scans
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 Schauer & Nuchter, 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 |
|---|---|---|
| peopleremover, voxel 17.5 cm本方法原文提出 | 0.95 fraction | (Schauer & Nuchter, 2018, Sec. IX) |
Lim et al., 2021 · Table III 本方法 1 筆
指標Runtime/iteration
資料集與序列SemanticKITTI · 01 (frames 150-250)
表格設定(擷取紀錄原文):Runtime per iteration of each dynamic-removal method on SemanticKITTI sequence 01; hardware not reported (Lim et al., 2021, Table III)
Runtime/iteration,SemanticKITTI · 01 (frames 150-250)
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。
- 未報告(沒有數值,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., 2021 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., 2021, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| OctoMap | 1.077 s | (Lim et al., 2021, Table III) |
| Peopleremover本方法 | 無數值未報告註記(擷取紀錄):other: printed as '1,000' s; decimal separator ambiguous | (Lim et al., 2021, Table III) |
| Removert | 0.8307 s | (Lim et al., 2021, Table III) |
| ERASOR (Ours)原文提出 | 0.0732 s | (Lim et al., 2021, Table III) |
來源
Schauer & Nuchter, 2018
(2018)The Peopleremover—Removing Dynamic Objects From 3-D Point Cloud Data by Traversing a Voxel Occupancy GridIEEE Robotics and Automation Letters, 3(3), pp. 1679-1686
同儕審查已出版已讀全文近十年