DeLORA
DeLORA 以自監督(self-supervised)方式訓練 LiDAR 里程計網路:推論時只輸入由原始掃描投影成的球面距離影像,網路直接輸出相鄰兩幀的相對位姿;訓練時以 KD-tree 在三維空間尋找對應點,計算點到平面與平面到平面的幾何損失,因此不需要真值位姿或標註資料。法向量在訓練前以主成分分析(PCA)離線計算,只隨預測旋轉而轉動,使梯度不必穿過法向量計算。作者在足式機器人 ANYmal(建物地下室長廊)、履帶機器人(DARPA SubT Urban Circuit 資料)與 KITTI 上測試,並把位姿接到 LOAM 建圖模組產生地圖。
本頁內容
A self-supervised LiDAR odometry network trained with point-to-plane and plane-to-plane losses on 3D nearest-neighbour correspondences, needing no ground-truth poses, and tested on legged, tracked and wheeled platforms.
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 | Velodyne VLP-16 Puck Lite (ANYmal, Sec. IV-A)、Ouster OS1-64 (DARPA SubT Urban, Sec. IV-B)、KITTI odometry LiDAR (sensor model not named in the paper, Sec. IV-C) |
|---|---|
| 原文測試平台 | legged (ANYmal)、tracked UGV (iRobot PackBot Explorer, DARPA SubT Urban Circuit dataset)、vehicle (KITTI) |
| 狀態估計 | self-supervised CNN (ResNet-like blocks) regressing relative 6-DoF pose (translation + quaternion) from two spherical range images; poses optionally passed to the LOAM mapping module (Sec. III-B, IV-A) |
| 資料關聯 | training only: 3D nearest-neighbour correspondences via KD-tree, with point-to-plane and plane-to-plane losses on PCA normals precomputed offline; inference uses the range image (x, y, z, range) only (Sec. III-C, III-D) |
| 時間表示 | discrete poses (scan-to-scan) |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | none inside the method; maps in the experiments are built by the LOAM mapping module fed with DeLORA poses (Sec. IV-A, IV-B) |
| 先驗資訊 | none at inference; training needs unlabelled scans from the target sensor or domain, no ground-truth poses (abstract, Sec. III) |
| 可輸出幾何 | 6-DoF relative poses; point-cloud maps only when combined with LOAM mapping (Figs. 3-4) |
| 計算需求 | about 48 ms per prediction on an i7-8565U laptop CPU and 13 ms on a GeForce MX250 laptop GPU, with about 32,000 points, H = 16, W = 720 (Sec. IV-A) |
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 Puck Lite歸入:Velodyne VLP-16 | 方法輸入 | 未標示 | mounted upright for the training missions and upside down for the test mission; range image H = 16, W = 720, about 32,000 points per scan | (Nubert et al., 2021, Sec. IV-A) |
| LiDAR | Ouster OS1-64 | 資料集感測器 | DARPA SubT Challenge Urban Circuit | carried by a tracked robot at Satsop Business Park; Alpha course used for training, Beta course for testing | (Nubert et al., 2021, Sec. IV-B) |
| 載具平台 | ANYmal quadrupedal robot | 方法輸入 | 未標示 | learning-based locomotion controller; autonomous exploration missions of about 250 m on average in the ETH Zurich CLA basement | (Nubert et al., 2021, Sec. IV-A; Fig. 1) |
| 載具平台 | iRobot PackBot Explorer (tracked robot) | 資料集感測器 | DARPA SubT Challenge Urban Circuit | tracked robot capable of fast in-spot yaw rotations | (Nubert et al., 2021, Sec. IV-B) |
| 運算硬體 | i7-8565U low-power laptop CPU | 執行運算平台 | 未標示 | about 48 ms per prediction | (Nubert et al., 2021, Sec. IV-A) |
| 運算硬體 | GeForce MX250 laptop GPU | 執行運算平台 | 未標示 | about 13 ms per prediction | (Nubert et al., 2021, Sec. IV-A) |
作者報告的優勢與限制
優勢
- no labelled or ground-truth data needed for training (abstract, Sec. III)
- runs in real time on a mobile-class CPU (Sec. IV-A)
- test mission used an upside-down LiDAR mounting while training used an upright mounting, and still produced a consistent map (Sec. IV-A, Fig. 4)
- trained on the SubT Alpha course and tested on the unseen Beta course (Sec. IV-B)
限制
- odometry-only errors on unseen KITTI sequences 09 and 10 (6.05% and 6.44%) are about 3.6 to 3.9 times those after LOAM scan-to-map refinement (1.54% and 1.78%) (derived from Table II; the authors state that refinement helps especially on the test set)
- ANYmal quantitative evaluation is relative to LOAM, because no external ground truth was available (Sec. IV-A, Table I)
- SubT evaluation against the ground-truth map is qualitative (Sec. IV-B, Fig. 3)
- IMU not used; the authors list multi-modal integration as future work (Sec. V)
- (inference) training data from each new sensor or domain is still needed; no construction-site test
營建工程相關證據
在 ETH Zürich 建物地下室的長廊(類隧道)以 ANYmal 足式機器人測試,並使用 DARPA SubT Urban Circuit 資料(Satsop Business Park 的核電廠設施);後者有地圖真值,但比較只作定性呈現。未在營建工地測試,也沒有以獨立量測評估點雲幾何精度。
原文驗證環境:已完工建築、基礎設施、公開基準
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 44 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。
Nubert et al., 2021 · Table I 本方法 12 筆
表格設定(擷取紀錄原文):ANYmal test mission (LiDAR mounted upside down); relative pose deviation of DeLORA poses combined with the LOAM mapping module against the open-source LOAM implementation, which serves as reference because no external ground truth was available (Nubert et al., 2021, Table I)
t_rel [%] relative translation deviation vs LOAM,ANYmal CLA basement (own data) · test mission, segment length 5 m
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 Nubert et al., 2021 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 |
|---|---|---|
| Ours with mapping (DeLORA + LOAM mapping module)本方法原文提出 | 0.345% | (Nubert et al., 2021, Table I) |
Nubert et al., 2021 · Table II 本方法 12 筆
表格設定(擷取紀錄原文):KITTI odometry, errors over all subsequences of 100 to 800 m; DeLORA trained self-supervised on 00-08, tested on 09 and 10; only the 00-08 mean of LO-Net and Velas et al. was adapted by the authors because those were trained on 00-06 (Nubert et al., 2021, Table II)
t_rel [%],KITTI odometry · Training 00-08 (mean)
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., 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(Nubert et al., 2021, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Ours本方法原文提出 | 3% | (Nubert et al., 2021, Table II) |
| Ours+Map (LOAM scan-to-map refinement)本方法原文提出 | 1.78% | (Nubert et al., 2021, Table II) |
| DeepLO [24] | 3.68% | (Nubert et al., 2021, Table II) |
| LO-Net [21] | 1.27% | (Nubert et al., 2021, Table II) |
| Velas et al. [20] | 2.94% | (Nubert et al., 2021, Table II) |
| UnDeepVO [8] | 4.54% | (Nubert et al., 2021, Table II) |
| SfMLearner [22] | 28.52% | (Nubert et al., 2021, Table II) |
| Zhu et al. [23] | 5.72% | (Nubert et al., 2021, Table II) |
| LO-Net+Map | 0.81% | (Nubert et al., 2021, Table II) |
| SUMA [13] | 3.06% | (Nubert et al., 2021, Table II) |
| LOAM [3] | 1.26% | (Nubert et al., 2021, Table II) |
Nubert et al., 2021 · Table III 本方法 8 筆
表格設定(擷取紀錄原文):Loss ablation on KITTI, networks trained from scratch on 00-06 and tested on 07-10 (Nubert et al., 2021, Table III)
t_rel [%],KITTI odometry · Training 00-06
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., 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(Nubert et al., 2021, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| DeLORA loss variant: p2pl + pl2pl本方法原文提出 | 3.41% | (Nubert et al., 2021, Table III) |
| DeLORA loss variant: p2pl本方法原文提出 | 6.47% | (Nubert et al., 2021, Table III) |
Deng et al., 2023 · Table 3 本方法 4 筆
指標RMSE of ATE (SE(3) alignment)
表格設定(擷取紀錄原文):Odometry ATE RMSE with SE(3) alignment (Sec. 5.1); '-' means failed; unit not stated; DeLORA and PWC-LONet are pre-trained on KITTI (Deng et al., 2023, Table 3)
RMSE of ATE (SE(3) alignment),MaiCity · Mai00
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 Deng 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(Deng et al., 2023, Table 3)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| ICP [3] (point-to-point) | 1.9 | (Deng et al., 2023, Table 3) |
| GICP [31] | 1.24 | (Deng et al., 2023, Table 3) |
| Puma [36] | 0.25 | (Deng et al., 2023, Table 3) |
| SuMA [2] | 2.01 | (Deng et al., 2023, Table 3) |
| DeLORA [26]本方法 | 57.57 | (Deng et al., 2023, Table 3) |
| PWC-LONet [39] | 3.28 | (Deng et al., 2023, Table 3) |
| Ours原文提出 | 1.27 | (Deng et al., 2023, Table 3) |
其他比較組
來源
Nubert et al., 2021
(2021)Self-supervised Learning of LiDAR Odometry for Robotic Applications2021 IEEE International Conference on Robotics and Automation (ICRA), Xi'an, China, pp. 9601-9607
DOI 10.1109/icra48506.2021.9561063arXiv 2011.05418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:Self-supervised Learning of LiDAR Odometry for Robotic Applications (arXiv v1 2020-11-10, v2 2021-06-25) https://arxiv.org/abs/2011.05418
- 程式碼釋出:DeLORA: Self-supervised Deep LiDAR Odometry for Robotic Applications https://github.com/leggedrobotics/DeLORA
程式碼:https://github.com/leggedrobotics/DeLORA(授權:BSD-3-Clause (LICENSE file, copyright 2021 Julian Nubert, Robotic Systems Lab, ETH Zurich))。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。