supereight
supereight 提出以八元樹(octree)為空間索引的稠密體積 SLAM 框架。最底層以 8×8×8 體素區塊為單位,並以 Morton 編碼排序與逐層遮罩做免鎖的平行配置,再預先計算三線性內插的查詢順序,使八元樹在 CPU 上的融合與射線投射效率接近 InfiniTAM 的雜湊表。同一資料結構可存放 TSDF,也可存放機率佔據(occupancy)地圖;佔據地圖改寫 Loop 等人的 b-spline 雜訊模型,改用對數勝算累加、機率截斷與依時間遺忘的更新,使其適用於增量 SLAM,並明確表示已觀測的空區。追蹤採 KinectFusion 式只用深度的點對面 ICP,作者並示範佔據地圖可直接供 Informed RRT* 路徑規劃查詢。
本頁內容
Octree-based dense volumetric SLAM on the CPU: Morton-coded octree with 8^3 voxel-block leaves and lock-free parallel allocation that stores either a TSDF or a SLAM-ready probabilistic occupancy field (b-spline noise model, log-odds, clamping, time-windowed forgetting), with depth-only KinectFusion ICP tracking and direct use of the same map for RRT* planning.
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 | RGB-D camera, depth only (TUM RGB-D real sequences and ICL-NUIM synthetic sequences; sensor models not named) |
|---|---|
| 原文測試平台 | public RGB-D sequences (TUM RGB-D real sequences; capture platform not described in the paper)、simulation (ICL-NUIM living room; offline MAV path-planning queries on the map) |
| 狀態估計 | KinectFusion-style frame-to-model point-to-plane ICP solved by Gauss-Newton on depth only, against vertex and normal maps ray-cast from the octree map |
| 資料關聯 | Projective data association between the current depth frame and the ray-cast model prediction |
| 時間表示 | discrete poses |
| 去畸變 | 不適用 (depth camera) |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | Morton-coded octree whose leaves are 8 x 8 x 8 voxel blocks allocated in parallel from a memory pool; the field type is generic: TSDF, or log-odds occupancy using a quadratic b-spline depth-noise model (sigma proportional to range squared, 4 cm at 2 m), clamping to [0.03, 0.97] and a time-windowed forgetting update (tau = 5 s) |
| 先驗資訊 | none |
| 可輸出幾何 | TSDF or occupancy octree map at 1 cm finest resolution (surfaces as zero crossings), usable directly for sampling-based path planning |
| 計算需求 | CPU only; experiments in SLAMBench on a Skylake i7-6700HQ with 16 GB; 10 to 40 Hz on a modern quad-core CPU (Abstract); planning timings on an Intel Core i7-6600U at 2.60 GHz |
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 |
|---|---|---|---|---|---|
| 運算硬體 | Skylake i7-6700HQ CPU with 16GB of memory | 執行運算平台 | 未標示 | Ubuntu 16.10, frequency scaling disabled, GCC 5.4.1; SLAMBench framework | (Vespa et al., 2018, Sec. V) |
| 運算硬體 | Intel Core i7-6600U CPU at 2.60 GHz | 執行運算平台 | 未標示 | used for the path-planning timings (GCC 5.4.0) | (Vespa et al., 2018, Sec. V-C) |
| 其他 | high frequency motion capture system (model not reported) | 參考或真值量測 | TUM RGB-D | trajectory ground truth of the TUM RGB-D sequences | (Vespa et al., 2018, Sec. V-A) |
作者報告的優勢與限制
優勢
- Tracking accuracy on par with InfiniTAM on ICL-NUIM and TUM with 1 cm voxels and depth-only tracking; the occupancy pipeline was best on ICL LR 3 and TUM fr3 office (Table I)
- Octree TSDF fusion and ray-casting time comparable to or faster than InfiniTAM voxel hashing (Sec. V-B, Fig. 6)
- Memory 1.95 to 13.77% (TSDF) and 3.01 to 22.52% (occupancy) of a pre-allocated grid at the same resolution (Table II)
- Planning queries faster than OctoMap: 12.6 ms versus 17.7 ms mean to the first feasible RRT* path and 1.57 ms versus 2.06 ms for trajectory optimization (Tables III and IV)
- Runs at 10 to 40 Hz on a quad-core CPU without GPU (Abstract)
限制
- All compared systems lost track on TUM fr1 floor and fr1 plant; the TSDF pipeline reached 0.758 m ATE on ICL LR 3 (Table I)
- Occupancy mapping is more expensive than TSDF fusion because of b-spline sampling, log-odds updates and explicit free-space storage (Sec. V-B)
- Depth-only ICP tracking without loop closure; combined geometric and photometric tracking left to future work (Sec. V-A)
- Runtime comparisons are between different code bases, so part of the differences come from implementation (Sec. V-B)
- Planning is shown as query timings on the map, not in flight; drone integration is future work (Sec. V-C, VI)
營建工程相關證據
論文未在施工現場測試,評估為 TUM RGB-D 與 ICL-NUIM 室內序列。同一個八元樹地圖同時支援追蹤與路徑規劃,且明確區分已觀測空區與未知區域,對室內工地巡檢機器人的規劃有參考價值(推論)。corpus 中 Vizzo et al., 2022 比較的 SuperEight 結果數值,依該紀錄引自 Wang 等人的 LiDAR 擴充版本,並非本文原始的 RGB-D 版本。
原文驗證環境:公開基準、模擬
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 42 筆紀錄。
Vespa et al., 2018 · Table I 本方法 20 筆
指標ATE (m)
表格設定(擷取紀錄原文):ATE RMSE (Euclidean distance between ground-truth and estimated positions) on ICL-NUIM living room and TUM RGB-D; 1 cm finest voxels, depth-only tracking, same parameters throughout; 'x' = tracking failure (Vespa et al., 2018, Table I)
ATE (m),ICL-NUIM · ICL_LR_0
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 Vespa et al., 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(Vespa et al., 2018, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| TSDF (octree TSDF fusion, ours)本方法原文提出 | 0.0113 m | (Vespa et al., 2018, Table I) |
| OFusion (octree occupancy fusion, ours)本方法原文提出 | 0.0305 m | (Vespa et al., 2018, Table I) |
| InfiniTAM [16] (default depth-only tracker) | 0.3052 m | (Vespa et al., 2018, Table I) |
Vespa et al., 2018 · Table II 本方法 16 筆
指標relative memory consumption compared to a pre-allocated grid
表格設定(擷取紀錄原文):Memory of the octree maps relative to a statically pre-allocated grid covering the same area at the same resolution (as in KinectFusion) (Vespa et al., 2018, Table II)
relative memory consumption compared to a pre-allocated grid,ICL-NUIM · LR_0
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 Vespa et al., 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(Vespa et al., 2018, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| TSDF (octree TSDF fusion, ours)本方法原文提出 | 7.67% | (Vespa et al., 2018, Table II) |
| OFusion (octree occupancy fusion, ours)本方法原文提出 | 11.15% | (Vespa et al., 2018, Table II) |
Vespa et al., 2018 · Table III 本方法 3 筆
資料集與序列原文未報告 (map built by the SLAM system)
表格設定(擷取紀錄原文):Time to find the first feasible straight-line path with Informed RRT* (OMPL) for an obstructed 2.83 m start-goal distance, 1 cm map, averaged over 10,000 executions (Vespa et al., 2018, Table III)
planning time (time),原文未報告 (map built by the SLAM system)
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 Vespa et al., 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(Vespa et al., 2018, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| OFusion (octree occupancy map, ours)本方法原文提出硬體:Intel Core i7-6600U at 2.60 GHz | 12.6 ms | (Vespa et al., 2018, Table III) |
| Octomap硬體:Intel Core i7-6600U at 2.60 GHz | 17.7 ms | (Vespa et al., 2018, Table III) |
Vespa et al., 2018 · Table IV 本方法 3 筆
資料集與序列原文未報告 (map built by the SLAM system)
表格設定(擷取紀錄原文):Time for linear optimization of a collision-free polynomial trajectory from the RRT* plan, averaged over 1,000 executions (Vespa et al., 2018, Table IV)
planning time (time),原文未報告 (map built by the SLAM system)
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 Vespa et al., 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(Vespa et al., 2018, Table IV)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| OFusion (octree occupancy map, ours)本方法原文提出硬體:Intel Core i7-6600U at 2.60 GHz | 1.57 ms | (Vespa et al., 2018, Table IV) |
| Octomap硬體:Intel Core i7-6600U at 2.60 GHz | 2.06 ms | (Vespa et al., 2018, Table IV) |
來源
Vespa et al., 2018
(2018)Efficient Octree-Based Volumetric SLAM Supporting Signed-Distance and Occupancy MappingIEEE Robotics and Automation Letters, 3(2):1144-1151
DOI 10.1109/lra.2018.2792537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
- accepted manuscript:Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping (Imperial Spiral, EVespaRAL_final.pdf) http://hdl.handle.net/10044/1/55715
- 程式碼釋出:emanuelev/supereight https://github.com/emanuelev/supereight
程式碼:https://github.com/emanuelev/supereight(授權:BSD-3-Clause for the core library with some MIT-licensed files (stated in README; no LICENSE file))。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。