Learning-based localizability
本文以神經網路直接由單一 LiDAR 掃描預測掃描對掃描配準在六個自由度上是否可定位,不需先建立對應或求解配準最佳化即可提早偵測失效。網路只用模擬資料訓練,並取代 CompSLAM 中以特徵值門檻判斷退化的模組;在礦坑隧道、開闊混凝土場地與辦公室的現地測試中,以地圖與曲線定性展示同一網路不需重新調參,也能轉用到 Ouster OS0-128。
本頁內容
Neural network trained on simulation predicts scan-to-scan localizability from raw LiDAR data before registration.
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 (Velodyne VLP-16; Ouster OS0-128 for sensor-transfer test) |
|---|---|
| 原文測試平台 | legged (ANYmal C)、simulation |
| 狀態估計 | sparse 3D convolutional ResUNet feature extractor (MinkowskiEngine; 4000 sampled points, 0.2 m voxels) with global max pooling and a 5-layer MLP giving six sigmoid outputs; multi-label binary classification of localizability along x, y, z, roll, pitch and yaw trained with binary cross-entropy; per-direction probability thresholds chosen from a validation precision-recall curve |
| 資料關聯 | 不適用 (prediction before registration) |
| 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 | none |
| 地圖表示 | 不適用 |
| 先驗資訊 | trained only on simulated scans from 15 environments (CAD meshes and meshed cave scans) rendered in Gazebo with a simulated VLP-16; labels from Monte Carlo sampling of 200 child scans registered by point-to-plane ICP and thresholded at 0.1 m and 2 deg |
| 可輸出幾何 | 不適用 |
| 計算需求 | inference 28 ms per scan on CPU (Intel i7 10700F) and 13 ms on an Nvidia RTX3070 GPU in a ROS node without code optimization; training about 4 h on an Nvidia RTX3090 |
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 | Velodyne VLP-16 | 方法輸入 | 未標示 | 16 beams, 30 deg vertical field of view; its simulated model generated all training scans in Gazebo | (Nubert et al., 2022b, Sec. V-A2, VI-B, VI-C) |
| LiDAR | Ouster OS0-128 | 方法輸入 | 未標示 | 128 beams, 90 deg vertical field of view; unseen in training, used for the generalization test | (Nubert et al., 2022b, Sec. VI-C) |
| 地面雷射掃描儀(TLS) | Leica RTC360 | 參考或真值量測 | 未標示 | used with the BLK2GO to record the ground-truth map of the mine tunnel | (Nubert et al., 2022b, Sec. VI-A) |
| 行動掃描設備 | Leica BLK2GO | 參考或真值量測 | 未標示 | used with the RTC360 to record the ground-truth map of the mine tunnel | (Nubert et al., 2022b, Sec. VI-A) |
| 載具平台 | ANYmal-C | 方法輸入 | 未標示 | quadrupedal robot; kinematic leg odometry fused in CompSLAM along non-localizable directions | (Nubert et al., 2022b, Sec. I, VI-B) |
| 運算硬體 | Nvidia RTX3090歸入:NVIDIA RTX 3090 | 執行運算平台 | 未標示 | training in about 4 h | (Nubert et al., 2022b, Sec. V-B) |
| 運算硬體 | Intel i7 10700F | 執行運算平台 | 未標示 | CPU-only inference 28 ms | (Nubert et al., 2022b, Sec. V-B) |
| 運算硬體 | Nvidia RTX3070 | 執行運算平台 | 未標示 | GPU inference 13 ms | (Nubert et al., 2022b, Sec. V-B) |
作者報告的優勢與限制
優勢
- No environment-specific threshold tuning; tested on two sensor types without modification (abstract)
- On the simulated tunnel test set ResUNet reached F1 0.517 and accuracy 0.854 versus 0.214 and 0.809 for PointNet (Table IV)
- In the Seemühle mine tunnel an eigenvalue threshold of at least 110 was needed for the classical detector, while the network flagged y-axis degeneracy without tuning (Fig. 4)
- The same network worked with an Ouster OS0-128 although smallest-eigenvalue scales differed markedly from the VLP-16 (Sec. VI-C, Fig. 7)
限制
- Detection only; mitigation relies on the downstream fusion framework (CompSLAM with leg odometry) (Sec. VI-B)
- Precision and recall are relatively low for z, roll and pitch because the training set has few non-localizable examples in these directions; test-set precision 0.398 (Sec. VI-A, Table IV)
- Binary per-axis labels; predicting a full 6-DOF covariance for robots misaligned with the environment is left to future work (Sec. VII)
- Field detection results are shown qualitatively only, with no quantitative metrics on real data (reviewer observation, Sec. VI-B, VI-C)
營建工程相關證據
現地測試之一在瑞士 Rümlang,原文先稱其為鋪面工作場址(paved work site)上的開闊場地,後又稱為工作場址旁的開闊場地;機器人由棚架下出發,走上周圍沒有幾何特徵的大片混凝土地面,掃描對掃描配準在 x、y 與偏航方向退化。原文未稱其為施工中工地,因此不歸類為 real_construction_site。另兩處為 Seemühle 礦坑隧道,以及 ETH 兼含室內空間與屋頂露台的辦公室環境(arXiv v2 的 Sec. VI-B)。
原文驗證環境:模擬、地下或隧道、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 2 個比較組,合計 14 筆紀錄。
Nubert et al., 2022b · Table IV 本方法 12 筆
表格設定(擷取紀錄原文):Localizability classification averaged over six dimensions; both networks trained 60 epochs on the same simulated splits; test set simulated from the tunnel ground-truth mesh (Nubert et al., 2022b, Table IV)
Accuracy,simulated localizability dataset · Train
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 Nubert et al., 2022b 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(Nubert et al., 2022b, Table IV)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| ResUNet (proposed)本方法原文提出 | 0.998 fraction | (Nubert et al., 2022b, Table IV) |
| PointNet | 0.957 fraction | (Nubert et al., 2022b, Table IV) |
Nubert et al., 2022b · Text Sec. V-B 本方法 2 筆
指標inference time
資料集與序列不適用 · per scan
表格設定(擷取紀錄原文):Inference time of the PyTorch model inside a ROS node, no specific code optimization (Nubert et al., 2022b, Text Sec. V-B)
inference time,不適用 · per scan
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 Nubert et al., 2022b 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(Nubert et al., 2022b, Text Sec. V-B)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| ResUNet (proposed), CPU-only本方法原文提出硬體:Intel i7 10700F (CPU only) | 28 ms | (Nubert et al., 2022b, Sec. V-B) |
| ResUNet (proposed), GPU本方法原文提出硬體:Nvidia RTX3070 | 13 ms | (Nubert et al., 2022b, Sec. V-B) |
來源
Nubert et al., 2022b
(2022)Learning-based Localizability Estimation for Robust LiDAR Localization2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 17-24
DOI 10.1109/iros47612.2022.9982257arXiv 2203.05698
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:Learning-based Localizability Estimation for Robust LiDAR Localization https://arxiv.org/abs/2203.05698