LIC-Fusion
LIC-Fusion 在多狀態約束卡爾曼濾波器(MSCKF)架構中,緊密融合 IMU、稀疏視覺特徵,以及從光達掃描中擷取並追蹤的邊緣與平面特徵點。其特色是線上估計三種非同步感測器之間的空間外參與時間偏移,以因應低成本裝置的延遲與時鐘偏差。系統為純里程計,不維護全域地圖,也不使用迴圈。
本頁內容
An MSCKF-based LiDAR-inertial-camera odometry that fuses tracked LiDAR edge/plane features and sparse visual features with online spatio-temporal calibration of all sensors.
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)、IMU (Xsens MTi-300)、monochrome global-shutter camera、GNSS RTK (reference only) |
|---|---|
| 原文測試平台 | wheeled UGV、handheld |
| 狀態估計 | MSCKF whose state holds the IMU state, camera-IMU and LiDAR-IMU extrinsics with time offsets (IMU clock as reference), and sliding windows of IMU clones at camera and LiDAR times; standard EKF update after measurement compression |
| 資料關聯 | LiDAR edge (high curvature) and surf (low curvature) points from scan rings tracked from the current scan to the previous scan by KD-tree nearest neighbours, giving point-to-line and point-to-plane distances with propagated covariance and chi-square Mahalanobis gating; FAST visual features tracked by KLT, triangulated from camera clones and used after MSCKF nullspace projection; all residuals compressed by Givens-rotation thin QR |
| 時間表示 | discrete cloned poses; camera and LiDAR time offsets relative to the IMU clock estimated online |
| 去畸變 | Not described: the full text contains no LiDAR motion-distortion compensation step; edge and surf features are taken from raw scan rings |
| 迴圈閉合 | none (purely odometry, no global map; Sec. III) |
| 全域最佳化 | none |
| 地圖表示 | none (sliding window of cloned states; no global map maintained) |
| 先驗資訊 | offline extrinsic calibration refined online |
| 可輸出幾何 | trajectory only (no map product reported) |
| 計算需求 | 原文未報告 in the full text; the method is described as single-thread |
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 | self-collected indoor and outdoor sequences | 原文未報告 | (Zuo et al., 2019, Sec. III; Fig. 2) |
| 慣性量測單元(IMU) | Xsens MTi-300 AHRS IMU歸入:Xsens MTi-300 | 方法輸入 | self-collected indoor and outdoor sequences | 原文未報告 | (Zuo et al., 2019, Sec. III; Fig. 2) |
| GNSS 接收器 | RTK GPS | 參考或真值量測 | self-collected outdoor sequence | centimeter-level accuracy, used as outdoor ground truth | (Zuo et al., 2019, Sec. III-A) |
| 相機 | Blackfly BFLY-PGE-23S6M | 方法輸入 | self-collected indoor and outdoor sequences | monochrome global-shutter camera | (Zuo et al., 2019, Sec. III; Fig. 2) |
| 載具平台 | custom Ackermann robot platform | 方法輸入 | self-collected outdoor sequence | outdoor carrier of the sensor rig | (Zuo et al., 2019, Sec. III-A) |
| 載具平台 | handheld sensor rig (held at chest height) | 方法輸入 | self-collected indoor sequences | indoor carrying mode | (Zuo et al., 2019, Sec. III-B) |
作者報告的優勢與限制
優勢
- Online spatial and temporal calibration among IMU, camera and LiDAR (abstract
- Sec. II) | Lower average ATE than MSCKF VIO and LOAM on an 800 m outdoor sequence with RTK GPS reference (Table I) | Only method that stayed usable in the violently shaken Indoor-C sequence (1.55 vs 49.94 for MSCKF and 2.44 for LOAM) (Table II)
限制
- No global map and no loop closure (Sec. III) | Evaluation limited to one outdoor sequence and indoor start-end drift (Sec. III) | LOAM had lower start-end error on Indoor-A (0.66 vs 0.98) and Indoor-B (0.46 vs 1.04) (Table II) | No runtime figures reported
營建工程相關證據
原文未報告
原文驗證環境: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 27 筆紀錄。
Zuo et al., 2020 · Table VI 本方法 14 筆
表格設定(擷取紀錄原文):Averaged ATE of 5 runs on 6 Vicon-room sequences (cluttered room, Vicon ground truth), orientation (deg) and position (m); ATE computed following Zhang and Scaramuzza [23]; '-' = translational error above 20 m. The Average column is printed by the authors (for LIO-MAP and LIC-Fusion it averages only the sequences that did not fail). (Zuo et al., 2020, Table VI)
averaged ATE, orientation (deg),Vicon Room sequences (authors' data) · Seq 1 (42.62 m)
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 Zuo et al., 2020 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(Zuo et al., 2020, Table VI)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LIC-Fusion 2.0原文提出 | 2.537 deg | (Zuo et al., 2020, Table VI) |
| OpenVINS-IC | 2.625 deg | (Zuo et al., 2020, Table VI) |
| Proposed-LI | 2.333 deg | (Zuo et al., 2020, Table VI) |
| LOAM | 5.88 deg | (Zuo et al., 2020, Table VI) |
| LIO-MAP | 無數值失敗註記(擷取紀錄):failed | (Zuo et al., 2020, Table VI) |
| LIC-Fusion本方法 | 2.345 deg | (Zuo et al., 2020, Table VI) |
Zuo et al., 2020 · Table V 本方法 7 筆
資料集與序列Teaching Building sequences (authors' data) · Seq 1 (about 108 m)
表格設定(擷取紀錄原文):Averaged start-to-end drift of 5 runs on 7 teaching-building sequences (Zhejiang University) that start and end at the same position; each cell printed as a three-component vector in metres without axis labels; '-' = severe failure with final drift norm larger than 30 m. No scalar is printed, so value is null and the printed vector is kept in metric_as_written. (Zuo et al., 2020, Table V)
averaged start-to-end drift error vector as printed: (-0.740, 0.0401, 0.222) m,Teaching Building sequences (authors' data) · Seq 1 (about 108 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 Zuo et al., 2020 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 |
|---|---|---|
| LIC-Fusion本方法 | 無數值未報告註記(擷取紀錄):other: three-component vector printed, no scalar reported | (Zuo et al., 2020, Table V) |
Zuo et al., 2019 · Table II 本方法 4 筆
指標average trajectory start-end error
表格設定(擷取紀錄原文):Indoor handheld sequences at chest height, normal to low light, slow to aggressive motion; no ground truth; average start-end error after returning to the start; unit not printed in the table (sequence lengths given in m) (Zuo et al., 2019, Table II)
average trajectory start-end error,self-collected indoor sequences · Indoor-A (39m)
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 Zuo et al., 2019 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(Zuo et al., 2019, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| MSCKF | 0.99 | (Zuo et al., 2019, Table II) |
| LIC-Fusion本方法原文提出 | 0.98 | (Zuo et al., 2019, Table II) |
| LOAM | 0.66 | (Zuo et al., 2019, Table II) |
Zuo et al., 2019 · Table I 本方法 2 筆
資料集與序列self-collected outdoor sequence · outdoor (~800 m)
表格設定(擷取紀錄原文):Outdoor ~800 m, 4 min sequence on an Ackermann robot with RTK GPS ground truth (centimeter level); average over 6 runs of the average ATE and its 1-sigma; alignment for the ATE not stated (Fig. 4 MSE uses a best-fit transform); LOAM output benefits from implicit loop closure through its global map (Zuo et al., 2019, Table I)
average of average absolute trajectory errors (ATE),self-collected outdoor sequence · outdoor (~800 m)
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 Zuo et al., 2019 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(Zuo et al., 2019, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| MSCKF | 10.75 m | (Zuo et al., 2019, Table I) |
| LIC-Fusion本方法原文提出 | 4.06 m | (Zuo et al., 2019, Table I) |
| LOAM | 23.08 m | (Zuo et al., 2019, Table I) |
來源
Zuo et al., 2019
(2019)LIC-Fusion: LiDAR-Inertial-Camera Odometry2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 5848-5854
DOI 10.1109/iros40897.2019.8967746arXiv 1909.04102
同儕審查已出版已讀全文近十年
相關版本
- 預印本:LIC-Fusion arXiv (v1 2019-09-09, v2 2019-11-01) https://arxiv.org/abs/1909.04102