CLIC
CLIC 以分段三次 B 樣條表示連續時間軌跡,在固定時間長度的滑動視窗內做平滑:LiDAR 點到平面、原始 IMU、偏差與視覺重投影因子都在各自量測時刻取軌跡位姿,並推導解析雅可比矩陣、以邊緣化保留舊狀態的資訊,使連續時間方法可即時運行。框架可接入一至兩顆 LiDAR 與相機,並線上估計相機與 IMU 相對 LiDAR 的時間偏移。
本頁內容
Continuous-time fixed-lag smoothing on a split cubic B-spline with LiDAR point-to-plane, raw IMU and visual reprojection factors, analytic Jacobians and marginalization, supporting multiple LiDARs, a camera and online time-offset calibration in real time.
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 | one or two 3D LiDARs、IMU、monocular camera (optional) |
|---|---|
| 原文測試平台 | MAV (NTU VIRAL)、handheld rigs (Newer College, LVI-SAM handheld, Vicon Room)、Jackal UGV (LVI-SAM dataset)、electric car (YQ) |
| 狀態估計 | continuous-time fixed-lag smoothing over a split cubic B-spline trajectory (knot spacing 0.03 s) with a 0.12 s LiDAR-inertial temporal window and a 10-keyframe visual window; Levenberg-Marquardt in Ceres with analytic Jacobians; marginalization keeps prior information when the window slides (Secs. III-V) |
| 資料關聯 | LiDAR point-to-plane factors from planar features, raw IMU factors, bias factors and visual reprojection factors with inverse depth; LiDAR points are evaluated on the continuous-time trajectory at their own timestamps (Sec. V) |
| 時間表示 | continuous time (split B-spline); every LiDAR point, IMU sample and image is evaluated at its timestamp; online camera and IMU time offsets with LiDAR as the base clock (Sec. V) |
| 去畸變 | implicit: LiDAR points are associated with trajectory poses at their own timestamps on the continuous-time trajectory |
| 迴圈閉合 | yes, Euclidean distance based detection with the two-stage continuous-time trajectory correction of CLINS (Sec. V) |
| 全域最佳化 | two-stage continuous-time loop closure correction after loop detection (Sec. V) |
| 地圖表示 | LiDAR feature map plus visual landmarks (Fig. 10) |
| 先驗資訊 | sensor extrinsics; time offsets and extrinsics can be calibrated online (Table V 'w/ calib' variant) |
| 可輸出幾何 | continuous-time trajectory and LiDAR map |
| 計算需求 | desktop PC with Intel i7-7700K and 32 GB RAM; on eee_01 (397 s) total processing 217.82 s for CLIO and 294.57 s for CLIC, faster than real time (Table VIII) |
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 | 16-beam Ouster (two units) | 方法輸入 | NTU VIRAL | 10 Hz | (Lv et al., 2023, Sec. VI-A) |
| LiDAR | 64-beam Ouster with internal IMU | 方法輸入 | Newer College Dataset | LiDAR 10 Hz; internal IMU 100 Hz; stereo camera 30 Hz; handheld device | (Lv et al., 2023, Sec. VI-A) |
| LiDAR | 16-beam LiDAR | 方法輸入 | LVI-SAM dataset | 10 Hz; camera 20 Hz; IMU 500 Hz; handheld and Jackal platforms | (Lv et al., 2023, Sec. VI-A) |
| LiDAR | 16-beam LiDAR (YQ and Vicon Room rig) | 方法輸入 | CLIC YQ and Vicon Room datasets (authors) | 10 Hz; camera 20 Hz; IMU 400 Hz; rig mounted on an electric car (YQ) or handheld (Vicon Room) | (Lv et al., 2023, Sec. VI-A; Fig. 6) |
| 慣性量測單元(IMU) | VN100 | 方法輸入 | NTU VIRAL | 385 Hz | (Lv et al., 2023, Sec. VI-A) |
| GNSS 接收器 | GPS (YQ ground truth) | 參考或真值量測 | CLIC YQ dataset (authors) | GPS measurements provide ground truth for the outdoor YQ dataset | (Lv et al., 2023, Sec. VI-A) |
| 相機 | monocular cameras (two) | 方法輸入 | NTU VIRAL | 10 Hz | (Lv et al., 2023, Sec. VI-A) |
| 運算硬體 | desktop PC with Intel i7-7700K | 執行運算平台 | 未標示 | 32 GB RAM | (Lv et al., 2023, Sec. VI-E) |
| 其他 | motion capture system | 參考或真值量測 | CLIC Vicon Room dataset (authors) | ground truth for the indoor Vicon Room dataset | (Lv et al., 2023, Sec. VI-A) |
作者報告的優勢與限制
優勢
- On NTU VIRAL, CLIO and CLIC average APE RMSE 0.034 and 0.035 m with one LiDAR versus 0.096 m for LIO-SAM; CLIO2 and CLIC2 0.034 m with two LiDARs (Table III)
- On Newer College NCD06 with vigorous handheld shaking, CLIO APE 0.091 m versus 0.272 m for LIO-SAM (Table IV)
- On the LVI-SAM handheld sequence, CLIC without loop closure 2.56 m versus 7.87 m for LVI-SAM, where LiDAR-only methods failed (Table V)
- Time offset converges within about 3 s in most trials with final mean -2.0 ms and std 2.7 ms (Sec. VI-D)
- Much faster than the earlier continuous-time CLINS (1601.86 s) on eee_01 (Table VIII)
限制
- CLIO fails on the LVI-SAM handheld sequence in open areas without the camera (Table V)
- On the Vicon Room dataset LIC-Fusion 2.0 is more accurate than CLIC on four of six sequences (Table VII)
- Several baseline values on NTU VIRAL are copied from the VIRAL SLAM preprint rather than re-run (Table III footnote)
營建工程相關證據
CLIC 的連續時間軌跡讓每個 LiDAR 點都在自己的時刻取位姿,等同內建點雲去畸變,對手持或無人機劇烈晃動時的點雲品質特別有幫助(推論),NCD06 手持劇烈晃動序列的結果支持這一點。測試包含校園建物、室內動作捕捉房與空曠植生區,但沒有施工現場或點雲精度評估。
原文驗證環境:公開基準、獨立參考量測、受控實驗
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 5 個比較組,合計 68 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 1 組列在最後,並連到性能比較頁。
Lv et al., 2023 · Table III 本方法 36 筆
指標APE (RMSE, meter)
表格設定(擷取紀錄原文):NTU VIRAL dataset (MAV, indoor and outdoor); APE RMSE in metres; sensors L = LiDAR, I = IMU, C = camera, L2 = two LiDARs; rows marked (2) are results quoted from [51] (VIRAL SLAM preprint) and their loop-closure setting is not stated; CLINS, CLIO and CLIC variants run without loop closure (Lv et al., 2023, Table III)
APE (RMSE, meter),NTU VIRAL · eee_01 (237 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 Lv 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(Lv et al., 2023, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LIO-SAM(2) [L, I] | 0.075 m | (Lv et al., 2023, Table III; Sec. VI) |
| MILIOM (horz. LiDAR)(2) [L, I] | 0.104 m | (Lv et al., 2023, Table III; Sec. VI) |
| VIRAL (horz. LiDAR)(2) [L, I] | 0.064 m | (Lv et al., 2023, Table III; Sec. VI) |
| CLINS (w/o loop) [L, I] | 0.059 m | (Lv et al., 2023, Table III; Sec. VI) |
| CLIO (w/o loop) [L, I]本方法原文提出 | 0.03 m | (Lv et al., 2023, Table III; Sec. VI) |
| CLIC (w/o loop) [L, I, C]本方法原文提出 | 0.03 m | (Lv et al., 2023, Table III; Sec. VI) |
| MILIOM (2 LiDARs)(2) [L2, I] | 0.067 m | (Lv et al., 2023, Table III; Sec. VI) |
| VIRAL (2 LiDARs)(2) [L2, I, C] | 0.06 m | (Lv et al., 2023, Table III; Sec. VI) |
| CLIO2 (w/o loop) [L2, I]本方法原文提出 | 0.04 m | (Lv et al., 2023, Table III; Sec. VI) |
| CLIC2 (w/o loop) [L2, I, C]本方法原文提出 | 0.038 m | (Lv et al., 2023, Table III; Sec. VI) |
Lv et al., 2023 · Table V 本方法 10 筆
指標APE (RMSE, meter)
表格設定(擷取紀錄原文):LVI-SAM dataset (handheld and Jackal, outdoor open vegetated and geometrically degenerate areas; 16-beam LiDAR 10 Hz, camera 20 Hz, IMU 500 Hz); APE RMSE in metres (Lv et al., 2023, Table V)
APE (RMSE, meter),LVI-SAM dataset · Handheld (1642 s)
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 Lv 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(Lv et al., 2023, Table V)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LIO-SAM (w/o loop) [L, I] | 53.62 m | (Lv et al., 2023, Table V; Sec. VI-C) |
| CLIO (w/o loop) [L, I]本方法原文提出 | 無數值失敗註記(擷取紀錄):failure ('fail' in Table V) | (Lv et al., 2023, Table V; Sec. VI-C) |
| LVI-SAM (w/o loop) [L, I, C] | 7.87 m | (Lv et al., 2023, Table V; Sec. VI-C) |
| CLIC (w/o loop) [L, I, C]本方法原文提出 | 2.56 m | (Lv et al., 2023, Table V; Sec. VI-C) |
| LIO-SAM (w/ loop) [L, I] | 無數值失敗註記(擷取紀錄):failure ('fail' in Table V) | (Lv et al., 2023, Table V; Sec. VI-C) |
| CLIO (w/ loop) [L, I]本方法原文提出 | 無數值失敗註記(擷取紀錄):failure ('fail' in Table V) | (Lv et al., 2023, Table V; Sec. VI-C) |
| LVI-SAM (w/ loop) [L, I, C] | 0.83 m | (Lv et al., 2023, Table V; Sec. VI-C) |
| CLIC (w/ loop) [L, I, C]本方法原文提出 | 0.65 m | (Lv et al., 2023, Table V; Sec. VI-C) |
| CLIC (w/ loop, w/ calib) [L, I, C]本方法原文提出 | 0.56 m | (Lv et al., 2023, Table V; Sec. VI-C) |
Lv et al., 2023 · Table VIII 本方法 10 筆
資料集與序列NTU VIRAL · eee_01 (397 s)
表格設定(擷取紀錄原文):Time consumption (seconds) of main modules over the whole eee_01 sequence (397 s) of NTU VIRAL on an Intel i7-7700K desktop with 32 GB RAM (Lv et al., 2023, Table VIII)
Update Local Map time over the sequence,NTU VIRAL · eee_01 (397 s)
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 Lv 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(Lv et al., 2023, Table VIII)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| CLINS硬體:desktop PC, Intel i7-7700K, 32 GB RAM | 17.39 s | (Lv et al., 2023, Table VIII; Sec. VI-E) |
| CLIO本方法原文提出硬體:desktop PC, Intel i7-7700K, 32 GB RAM | 11.56 s | (Lv et al., 2023, Table VIII; Sec. VI-E) |
| CLIC本方法原文提出硬體:desktop PC, Intel i7-7700K, 32 GB RAM | 11.52 s | (Lv et al., 2023, Table VIII; Sec. VI-E) |
Lv et al., 2023 · Table IV 本方法 6 筆
指標APE (RMSE, meter)
表格設定(擷取紀錄原文):Newer College Dataset (handheld, 64-beam Ouster with internal IMU); APE RMSE in metres; LiDAR-IMU methods only (Lv et al., 2023, Table IV)
APE (RMSE, meter),Newer College Dataset · NCD_01 (1530 s / 1609 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 Lv 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(Lv et al., 2023, Table IV)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LIO-SAM (w/o loop) | 1.66 m | (Lv et al., 2023, Table IV; Sec. VI-B) |
| CLIO (w/o loop)本方法原文提出 | 0.792 m | (Lv et al., 2023, Table IV; Sec. VI-B) |
| LIO-SAM (w/ loop) | 0.544 m | (Lv et al., 2023, Table IV; Sec. VI-B) |
| CLIO (w/ loop)本方法原文提出 | 0.408 m | (Lv et al., 2023, Table IV; Sec. VI-B) |
其他比較組
列出其餘 1 個比較組
來源
Lv et al., 2023
(2023)Continuous-Time Fixed-Lag Smoothing for LiDAR-Inertial-Camera SLAMIEEE/ASME Transactions on Mechatronics, 28(4), pp. 2259-2270
DOI 10.1109/tmech.2023.3241398arXiv 2302.07456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:CLIC arXiv v1 https://arxiv.org/abs/2302.07456
- 程式碼釋出:APRIL-ZJU/clic (GPL-3.0 per README) https://github.com/APRIL-ZJU/clic
程式碼:https://github.com/APRIL-ZJU/clic(授權:GPL-3.0 (README))。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。