RF-LIO
RF-LIO 以 LIO-SAM 為基礎,處理大量移動物體時「先要準確位姿才能移除動態點、但動態點又破壞配準」的循環問題。新關鍵影格到達時先不做掃描配準,而是以 IMU 預積分取得初始位姿,並依預測的平移與旋轉誤差決定距離影像的角解析度;將目前掃描與周邊特徵子地圖投影成同解析度距離影像,依可見性差異移除子地圖中的動態點,再做 LOAM 特徵配準。若以邊緣點距離計算的收斂分數未達門檻,就以新解析度重複移除與配準;收斂並完成圖最佳化後,再以細解析度移除目前關鍵影格殘留的動態點。
本頁內容
LIO-SAM-based LIO for highly dynamic scenes that removes moving points before scan matching: the IMU-predicted pose error sets the resolution of scan and submap range images, visibility differences flag dynamic points, and removal and LOAM-feature matching are repeated at new resolutions until an edge-distance convergence score is met.
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 at 10 Hz (model not named)、IMU at 400 Hz (model not named) |
|---|---|
| 原文測試平台 | 原文未報告 (self-collected Urban, Campus and Suburban datasets and UrbanLoco CA sequences; carrier not described in the paper) |
| 狀態估計 | LIO-SAM factor graph in GTSAM (IMU preintegration, LiDAR odometry and loop closure factors) with an added removal-first loop: IMU prior, dynamic-point removal, scan matching, convergence check and repeated removal at a new resolution (Sec. III-A, IV-A) |
| 資料關聯 | LOAM edge and planar features with nearest-neighbour point-to-line and point-to-plane distances to a feature submap, after removing submap points flagged as dynamic (Sec. III-E) |
| 時間表示 | discrete keyframes with IMU preintegration (Sec. III-B) |
| 去畸變 | IMU-based motion compensation of each scan as in LIO-SAM (Sec. III-A, Fig. 2) |
| 迴圈閉合 | yes; same loop detection as LIO-SAM (Sec. IV-A) |
| 全域最佳化 | factor-graph optimization with GTSAM as in LIO-SAM (Sec. IV-A) |
| 地圖表示 | keyframe-based edge and planar feature map from which moving points are removed; global point map without ghost tracks (Fig. 5) |
| 先驗資訊 | none |
| 可輸出幾何 | trajectory and a static point cloud map with moving-object points removed |
| 計算需求 | per-scan runtime of RF-LIO (FA) 61 to 121 ms on an Intel i7-10700K CPU; below 100 ms on low and medium dynamic data except Suburban (112 ms) (Table VI, 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 | 3D LiDAR (model not stated) | 方法輸入 | 未標示 | LiDAR scans at 10 Hz (Fig. 2) | (Qian et al., 2021, Fig. 2; Sec. IV-A) |
| 慣性量測單元(IMU) | IMU (model not stated) | 方法輸入 | 未標示 | IMU data at 400 Hz (Fig. 2) | (Qian et al., 2021, Fig. 2; Sec. III-B; Sec. IV-A) |
| GNSS 接收器 | GPS (receiver model not stated) | 參考或真值量測 | self-collected Urban, Campus, Suburban and UrbanLoco | used only as ground truth | (Qian et al., 2021, Sec. IV-A) |
| 運算硬體 | Intel i7-10700k (as written)歸入:Intel i7-10700k | 執行運算平台 | 未標示 | ROS on Ubuntu Linux | (Qian et al., 2021, Sec. IV-A; Table VI) |
作者報告的優勢與限制
優勢
- Average moving-point removal rate of 96.1% relative to LIO-SAM maps on three self-collected datasets (Table III)
- On UrbanLoco high-dynamic sequences, ATE RMSE 15.89 and 12.17 m versus 62.43 and 36.98 m for LIO-SAM and 203.78 and 175.69 m for LOAM (Table V)
- Removal-first variants also improved ATE on low and medium dynamic data, e.g. Urban 6.46 m versus 10.21 m for LIO-SAM (Table IV)
- Removal before matching reduced per-scan runtime compared with removal after matching (Table VI)
- No training data or semantic labels required (Sec. V)
限制
- In very open scenes without far points behind moving objects, the visibility test cannot remove moving points (Sec. V)
- Not suitable when moving objects fully block the sensor field of view (Sec. V)
- Points close to the ground (below 0.5 m) and returns from beams parallel to the ground are not removed (Sec. IV-C)
- Runtime exceeds the 100 ms LiDAR period on Suburban and CARussianHill (112 and 121 ms) (Table VI)
- Ground truth is GPS only; the removal rate is measured relative to LIO-SAM's residual moving points rather than labelled ground truth (Sec. IV-A, IV-C)
營建工程相關證據
論文未在營建場域評估;自建 Urban 資料的路線雖經過施工區,但未單獨分析(Sec. IV-D)。其先移除動態點再配準的做法,概念上可對應工地上移動的工人、車輛與機具造成的重影與配準偏差,對需要靜態點雲地圖的施工記錄有參考價值;但在開闊或被大型機具遮擋視野的工地,作者指出的兩項限制可能直接出現(推論)。
原文驗證環境:公開基準、跨場域
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 38 筆紀錄。
Qian et al., 2021 · Table VI 本方法 15 筆
指標runtime for processing one scan
表格設定(擷取紀錄原文):Runtime of RF-LIO variants for processing one scan (Qian et al., 2021, Table VI)
runtime for processing one scan,self-collected datasets and UrbanLoco · Urban
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 Qian 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(Qian et al., 2021, Table VI)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| RF-LIO (After)本方法硬體:Intel i7-10700K | 86 ms | (Qian et al., 2021, Table VI) |
| RF-LIO (First)本方法原文提出硬體:Intel i7-10700K | 55 ms | (Qian et al., 2021, Table VI) |
| RF-LIO (FA)本方法原文提出硬體:Intel i7-10700K | 61 ms | (Qian et al., 2021, Table VI) |
Qian et al., 2021 · Table IV 本方法 9 筆
指標absolute trajectory RMSE
表格設定(擷取紀錄原文):Low and medium dynamic self-collected datasets; LiDAR and IMU only, GPS as ground truth; RF-LIO and LIO-SAM share feature extraction and loop closure (Qian et al., 2021, Table IV)
absolute trajectory RMSE,self-collected datasets · Urban (6390.33 m, low dynamic)
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 Qian 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(Qian et al., 2021, Table IV)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LOAM | 244.19 m | (Qian et al., 2021, Table IV) |
| LIO-SAM | 10.21 m | (Qian et al., 2021, Table IV) |
| RF-LIO (After)本方法 | 10.79 m | (Qian et al., 2021, Table IV) |
| RF-LIO (First)本方法原文提出 | 7.72 m | (Qian et al., 2021, Table IV) |
| RF-LIO (FA)本方法原文提出 | 6.46 m | (Qian et al., 2021, Table IV) |
Qian et al., 2021 · Table III 本方法 8 筆
表格設定(擷取紀錄原文):Residual moving-object points counted in the maps of LIO-SAM and RF-LIO (same feature extraction); removal rate relative to LIO-SAM (Qian et al., 2021, Table III)
residual moving object points in map,self-collected datasets · Urban
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 Qian 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(Qian et al., 2021, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LIO-SAM | 85890 points | (Qian et al., 2021, Table III) |
| RF-LIO本方法原文提出 | 1803 points | (Qian et al., 2021, Table III) |
Qian et al., 2021 · Table V 本方法 6 筆
指標absolute trajectory RMSE
表格設定(擷取紀錄原文):High dynamic UrbanLoco sequences with many moving objects; LiDAR and IMU only, GPS as ground truth (Qian et al., 2021, Table V)
absolute trajectory RMSE,UrbanLoco · CAMarketStreet (5690.98 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 Qian 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(Qian et al., 2021, Table V)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LOAM | 203.78 m | (Qian et al., 2021, Table V) |
| LIO-SAM | 62.43 m | (Qian et al., 2021, Table V) |
| RF-LIO (After)本方法 | 23.98 m | (Qian et al., 2021, Table V) |
| RF-LIO (First)本方法原文提出 | 15.83 m | (Qian et al., 2021, Table V) |
| RF-LIO (FA)本方法原文提出 | 15.89 m | (Qian et al., 2021, Table V) |
來源
Qian et al., 2021
(2021)RF-LIO: Removal-First Tightly-coupled Lidar Inertial Odometry in High Dynamic Environments2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 4421-4428
DOI 10.1109/iros51168.2021.9636624arXiv 2206.09463
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:arXiv 2206.09463 v1 (2022-06-19), posted after the conference https://arxiv.org/abs/2206.09463