Dynablox
Dynablox 延伸 Voxblox 的雜湊區塊體素地圖,在機器人運作中逐步估計「高信心自由空間」,並同時建模感測雜訊與稀疏性、狀態估計漂移及地圖不完整;落入高信心自由空間的點即判定為移動點,再以其為種子擴張叢集。方法不假設物體外觀或類別,可偵測搬運物品的人、擺動的門等多樣動態物。
本頁內容
Dynablox detects moving points online as those falling into conservatively estimated high-confidence free space in a volumetric map, accounting for noise, sparsity, drift and incomplete mapping.
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 | 未查證 |
| 狀態估計 | 不適用 (uses external state estimate, e.g., FAST-LIO2 in new sequences) |
| 資料關聯 | a voxel is occupied if its TSDF distance is below 1.5 voxel sizes or a current point falls in it, with a sparsity compensation of 2 frames; it becomes high-confidence free only after 5 frames unoccupied for itself and all observed neighbours; points in or next to free voxels seed clusters grown over connected voxels, clusters under 20 voxels discarded; slowly re-occupied voxels are reset after τr frames derived from the expected drift rate (Sec. IV-C, IV-D) |
| 時間表示 | 不適用 |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | 不適用 |
| 全域最佳化 | 不適用 |
| 地圖表示 | Voxblox hashed voxel blocks (16^3 voxels per block, voxel size 0.2 m) holding a TSDF; each voxel additionally stores the last occupied frame, the occupancy duration and a high-confidence free flag; voxels in dynamic clusters are overwritten rather than averaged at the next update (Sec. IV-A, IV-D) |
| 先驗資訊 | online state estimate with a drift-tolerance parameter |
| 可輸出幾何 | per-point dynamic labels during online mapping; static volumetric map |
| 計算需求 | real-time on a NUC with laptop-grade AMD 4800U CPU: pre-processing constant at 13.0 ms; total 58.1 ms per frame (17.2 FPS) at 20 m integration distance; TSDF integration and free-space update scale with the number of updated blocks, so unbounded range is slower in large open spaces such as the Station scenes (Sec. V, VII-D, Fig. 5) |
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 | OS1 64歸入:Ouster OS1-64 | 資料集感測器 | DOALS | high-range, high-resolution, 10 Hz; 8 sequences in 4 environments | (Schmid et al., 2023, Sec. VI) |
| LiDAR | Ouster OS0 | 方法輸入 | Dynablox newly recorded sequences | 128 beams (written 'OS0 128' in Sec. VI and '128-beam Ouster OS0' in Sec. IV-C), high-range, high-resolution, 90 deg vertical FoV, 10 Hz; state estimates from FAST-LIO2 | (Schmid et al., 2023, Sec. IV-C, Sec. VI) |
| 運算硬體 | NUC with laptop-grade AMD-4800U CPU | 執行運算平台 | 未標示 | also used in some of the authors' aerial and ground robots; all experiments | (Schmid et al., 2023, Sec. VI) |
作者報告的優勢與限制
優勢
- 86.0% IoU over all DOALS sequences with a 20 m range and 83.8% at full range (up to 172.7 m), above the learning-based baselines and below the offline Occupancy upper bound of 88% (Table I, Sec. VII-A)
- 58.1 ms per frame (17.2 FPS) at 20 m integration distance on an AMD 4800U NUC, about 39% (38.8%) above conventional TSDF mapping (Sec. VII-D)
- Recall largely independent of simulated drift and still 72% in the worst case; precision loss under drift mitigated by τr, whose correct setting improves performance by up to 88% (Sec. VII-C, Fig. 4)
- Class-agnostic: detected people carrying boxes, rolling cases or surfboards, rolling balls and swinging doors on stairs and across storeys (Sec. VII-B, Fig. 3)
限制
- Thin, sparsely measured objects, reflective surfaces such as windows, and strong occlusions cause failures (Sec. VII-F)
- Relies on dense mapping requiring sufficiently high data rates for the sensor speed (Sec. VII-F)
- Stated in follow-up work by theory-based analysis (not an experiment): cannot handle non-sequential data such as static survey scans (Duberg et al., 2024 Sec. V-B2)
營建工程相關證據
作者以多層建築與樓梯等複雜室內場景做定性示範(abstract);無工地測試。窗戶等反射面與細薄物體為已報告失效(Sec. VII-F),與施工中建築的玻璃、鷹架、防護網相關(推論)。
原文驗證環境:公開基準、已完工建築、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 8 個比較組,合計 46 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 4 組列在最後,並連到性能比較頁。
Duberg et al., 2024 · Table I 本方法 12 筆
表格設定(擷取紀錄原文):Point-wise dynamic point removal accuracy (%) following the DynamicMap benchmark protocol; Removert, ERASOR, OctoMap and DUFOMap evaluated offline, Dynablox and DUFOMap* online (each scan classified with the map built so far); DUFOMap uses the same parameters for all data (voxel 0.1 m, ds 0.2 m, dp 1), Removert and ERASOR per-dataset optimized parameters; KITTI labels and poses from SemanticKITTI (Duberg et al., 2024, Table I)
SA (static accuracy, share of static points correctly kept),KITTI (SemanticKITTI labels and poses) · 00 small town
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 Duberg et al., 2024 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(Duberg et al., 2024, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Removert [8] | 99.44% | (Duberg et al., 2024, Table I) |
| ERASOR [9] | 66.7% | (Duberg et al., 2024, Table I) |
| OctoMap [16] | 68.05% | (Duberg et al., 2024, Table I) |
| DUFOMap (Ours)原文提出 | 97.96% | (Duberg et al., 2024, Table I) |
| Dynablox [17]本方法 | 96.76% | (Duberg et al., 2024, Table I) |
| DUFOMap* (Ours, online)原文提出 | 98.37% | (Duberg et al., 2024, Table I) |
Schmid et al., 2023 · Table I 本方法 10 筆
指標Dynamic point detection IoU [%]
表格設定(擷取紀錄原文):DOALS (OS1 64 at 10 Hz, 8 sequences in 4 environments): IoU (%) between detected and annotated dynamic points, mean over the 10 annotated frames per sequence; full range up to 172.7 m unless marked (20m); Occupancy is offline with past and future scans (upper limit); DOALS-3DMiniNet trained on the other DOALS environments; 4DMOS, LMNet and MotionSeg3D pre-trained on KITTI, 'Refit' refits model statistics on DOALS; NUC with AMD 4800U (Schmid et al., 2023, Table I)
Dynamic point detection IoU [%],DOALS · Station
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。
- 失敗
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 Schmid et al., 2023 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(Schmid et al., 2023, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Occupancy [10] (Offline) | 91% | (Schmid et al., 2023, Table I) |
| DOALS-3DMiniNet [10,28] | 84% | (Schmid et al., 2023, Table I) |
| 4DMOS [14] | 38.8% | (Schmid et al., 2023, Table I) |
| LMNet [8] (Original) | 6% | (Schmid et al., 2023, Table I) |
| LMNet [8] (Refit) | 19.9% | (Schmid et al., 2023, Table I) |
| MotionSeg3D [9] | 無數值失敗註記(擷取紀錄):no meaningful result (marked x in table) | (Schmid et al., 2023, Table I) |
| Ours本方法原文提出 | 86.2% | (Schmid et al., 2023, Table I) |
| LC Free Space [22] (20m) | 48.7% | (Schmid et al., 2023, Table I) |
| Ours (20m)本方法原文提出 | 87.3% | (Schmid et al., 2023, Table I) |
Duberg et al., 2024 · Table III 本方法 9 筆
表格設定(擷取紀錄原文):Influence of the pose source on dynamic point removal, KITTI sequence 00: KITTI odometry ground-truth poses, SemanticKITTI poses estimated by SuMa, and KISS-ICP poses; SA, DA, AA in % (Duberg et al., 2024, Table III)
SA (static accuracy, share of static points correctly kept),KITTI (SemanticKITTI labels) · 00 (KITTI GT poses)
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 Duberg et al., 2024 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(Duberg et al., 2024, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Removert [8] | 99.18% | (Duberg et al., 2024, Table III) |
| ERASOR [9] | 63.83% | (Duberg et al., 2024, Table III) |
| Octomap [16] | 54.81% | (Duberg et al., 2024, Table III) |
| Dynablox [17]本方法 | 95.5% | (Duberg et al., 2024, Table III) |
| DUFOMap (Ours)原文提出 | 92.57% | (Duberg et al., 2024, Table III) |
Schmid et al., 2023 · Table II 本方法 7 筆
指標IoU [%] at 20 m
資料集與序列DOALS · All
表格設定(擷取紀錄原文):Ablation of the modelled components at 20 m range on DOALS (IoU %, all sequences) (Schmid et al., 2023, Table II)
IoU [%] at 20 m,DOALS · All
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 Schmid et al., 2023 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(Schmid et al., 2023, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Ours本方法原文提出 | 86% | (Schmid et al., 2023, Table II) |
| Dynablox (w/o Occupancy Cue)本方法原文提出 | 85.6% | (Schmid et al., 2023, Table II) |
| Dynablox (w/o TSDF Cue)本方法原文提出 | 23.6% | (Schmid et al., 2023, Table II) |
| Dynablox (w/o temporal window τw)本方法原文提出 | 11.6% | (Schmid et al., 2023, Table II) |
| Dynablox (w/o Sparsity Comp. τs)本方法原文提出 | 83.3% | (Schmid et al., 2023, Table II) |
| Dynablox (w/o Spatial Margin N)本方法原文提出 | 38.1% | (Schmid et al., 2023, Table II) |
| Dynablox (w/o Cluster Filter τc)本方法原文提出 | 85.3% | (Schmid et al., 2023, Table II) |
其他比較組
來源
Schmid et al., 2023
(2023)Dynablox: Real-Time Detection of Diverse Dynamic Objects in Complex EnvironmentsIEEE Robotics and Automation Letters, 8(10):6259-6266
DOI 10.1109/lra.2023.3305239arXiv 2304.10049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:ethz-asl/dynablox https://github.com/ethz-asl/dynablox
程式碼:https://github.com/ethz-asl/dynablox(授權:BSD-3-Clause (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。