SDV-LOAM
SDV-LOAM 把視覺與 LiDAR 分成前後兩個模組:視覺模組是半直接法深度增強視覺里程計,先以光度誤差直接估計位姿,再做帶傳播的點匹配與重投影修正,並以滑動視窗光束法平差最佳化,追蹤點的深度直接取自投影的 LiDAR 點;其位姿作為 LiDAR 模組的運動先驗。LiDAR 模組以 CT-ICP 為基礎,將 10 Hz 掃描重組成與 60 Hz 影像同步的片段,並依地面與垂直約束比例在 3 自由度與 6 自由度的掃描對地圖最佳化之間切換,以減少垂直方向漂移。
本頁內容
Cascaded visual-LiDAR odometry: a semi-direct depth-enhanced monocular VO (direct pose, point matching with propagation, windowed BA; LiDAR depth without interpolation) feeds a CT-ICP-based LiDAR odometry that uses sweep reconstruction to match the camera rate and an adaptive 3-DoF or 6-DoF sweep-to-map optimization.
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 | monocular grayscale camera、3D LiDAR (Velodyne HDL-64E on KITTI; VLP-16 on the authors' rig) |
|---|---|
| 原文測試平台 | KITTI, KITTI-360 and KITTI-CARLA vehicles、authors' camera-LiDAR rig on a vehicle (qualitative) |
| 狀態估計 | semi-direct depth-enhanced visual odometry built on DSO ideas: direct photometric pose estimate, then point matching with propagation, reprojection-based refinement and sliding-window bundle adjustment with marginalization; its pose is the motion prior for a CT-ICP-based LiDAR odometry with adaptive sweep-to-map optimization (Secs. V-VI) |
| 資料關聯 | visual tracking points take depth from projected LiDAR points without depth interpolation, plus extra high-gradient points without LiDAR depth; LiDAR sweep-to-map point-to-plane registration against a voxel map (20 points per voxel) switching between 3-DoF and 6-DoF optimization by the ratio of ground to vertical constraints (threshold 0.8) (Secs. V-VI) |
| 時間表示 | sweep reconstruction splits and recombines 10 Hz LiDAR sweeps into 60 Hz segments aligned with the camera frames (Sec. VI; Fig. 5); the LiDAR module is built on CT-ICP, a continuous-time registration method |
| 去畸變 | not described as a separate step; the LiDAR module is based on CT-ICP (inference: intra-sweep motion is handled by its continuous-time model) |
| 迴圈閉合 | no |
| 全域最佳化 | none |
| 地圖表示 | voxel map of LiDAR points for sweep-to-map registration; global point cloud map output (Fig. 8) |
| 先驗資訊 | camera-LiDAR extrinsic calibration (Autoware calibration toolkit for the authors' rig) |
| 可輸出幾何 | vehicle trajectory and global LiDAR point cloud map |
| 計算需求 | laptop with Intel i7-11700 and 16 GB RAM; visual odometry about 0.06 s per frame; both modules reach about 20 Hz in practice (Sec. VII) |
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 | VLP-16 (introduction says VLP-16E)歸入:Velodyne VLP-16 | 方法輸入 | SDV-LOAM own platform (qualitative) | 16-beam LiDAR at 10 Hz | (Yuan et al., 2023a, Sec. VII; Fig. 6) |
| LiDAR | Velodyne HDL-64E | 資料集感測器 | KITTI odometry | KITTI LiDAR | (Yuan et al., 2023a, Sec. VII) |
| GNSS 接收器 | KITTI high accuracy GPS/INS | 參考或真值量測 | KITTI odometry | ground truth for KITTI | (Yuan et al., 2023a, Sec. VII) |
| 相機 | FL3-U3-13E4M-C (text also names FL3-FW-14S3M-C) | 方法輸入 | SDV-LOAM own platform (qualitative) | grayscale images 1280 x 1040 at 60 Hz | (Yuan et al., 2023a, Sec. VII; Fig. 6) |
| 雙目相機 | KITTI color and monochrome stereo cameras (model not named; only the left image is used) | 資料集感測器 | KITTI odometry | 10 Hz; SDV-LOAM takes the LiDAR point cloud and the left image of the stereo camera as input | (Yuan et al., 2023a, Sec. VII-A) |
| 運算硬體 | laptop with Intel i7-11700 | 執行運算平台 | 未標示 | 16 GB RAM | (Yuan et al., 2023a, Sec. VII) |
作者報告的優勢與限制
優勢
- Visual module KITTI 00-10 average RTE 0.72 % versus 1.16 % for DEMO and 0.94 % for Huang et al. (Table V)
- Full system KITTI 00-10 average 0.47 % versus 0.49 % for CT-ICP+ and 0.52 % for MULLS+ (Table VI)
- KITTI online test set 0.60 % translational drift, reported as 8th at the time of writing (Sec. VII)
- The adaptive sweep-to-map optimization was applied to six open-source LiDAR odometry systems to reduce vertical drift (Sec. VII ablations)
- Code released (GPL-2.0)
限制
- No loop closure or global optimization
- Own-rig experiments are qualitative (trajectory overlaid on a map; no ground truth)
- Baseline values in Table V come from the original publications because some codes could not be compiled or were not released (Sec. VII)
- On the KITTI test set V-LOAM (0.54 %) remains lower than SDV-LOAM (0.60 %) (Table VI)
營建工程相關證據
SDV-LOAM 屬於相機與 LiDAR 串接的里程計,評估全在 KITTI 系列車載資料與自建車載設備的定性結果,沒有室內或施工現場測試。其自適應 3 與 6 自由度最佳化是為了處理地面主導、垂直約束弱的場景,對開闊工地的車載或手推式掃描有參考價值(推論),但缺乏迴圈閉合,長距離點雲的全域一致性需另行處理。
原文驗證環境:公開基準、模擬
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 26 筆紀錄。
Yuan et al., 2023a · Table V 本方法 13 筆
指標Relative translational error (RTE)
表格設定(擷取紀錄原文):KITTI odometry; relative translational error (%) of LiDAR-assisted depth-enhanced visual odometry; baseline values from the original publications; LIMO* uses semantic information; '-' cells not stored (Yuan et al., 2023a, Table V)
Relative translational error (RTE),KITTI odometry · 00
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 Yuan et al., 2023a 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(Yuan et al., 2023a, Table V)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| DEMO | 1.05% | (Yuan et al., 2023a, Table V; Sec. VII) |
| LIMO* | 1.12% | (Yuan et al., 2023a, Table V; Sec. VII) |
| Huang et al. | 0.99% | (Yuan et al., 2023a, Table V; Sec. VII) |
| DVL-SLAM | 0.93% | (Yuan et al., 2023a, Table V; Sec. VII) |
| Our VO module本方法原文提出 | 0.67% | (Yuan et al., 2023a, Table V; Sec. VII) |
Yuan et al., 2023a · Table VI (KITTI part) 本方法 13 筆
指標Relative translational error (RTE)
表格設定(擷取紀錄原文):KITTI odometry; relative translational error (%) of visual-LiDAR odometry; '+' = open-source LiDAR odometry modified by the authors to use the SDV-LOAM visual module as motion prior; V-LOAM only has test-set results; '-' cells not stored (Yuan et al., 2023a, Table VI (KITTI part))
Relative translational error (RTE),KITTI odometry · 00
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 Yuan et al., 2023a 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(Yuan et al., 2023a, Table VI (KITTI part))
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| A-LOAM+ | 0.88% | (Yuan et al., 2023a, Table VI; Sec. VII) |
| LeGO-LOAM+ | 1.25% | (Yuan et al., 2023a, Table VI; Sec. VII) |
| Fast-LOAM+ | 0.76% | (Yuan et al., 2023a, Table VI; Sec. VII) |
| ISC-LOAM+ | 1.03% | (Yuan et al., 2023a, Table VI; Sec. VII) |
| MULLS+ | 0.52% | (Yuan et al., 2023a, Table VI; Sec. VII) |
| CT-ICP+ | 0.49% | (Yuan et al., 2023a, Table VI; Sec. VII) |
| Ours (SDV-LOAM)本方法原文提出 | 0.5% | (Yuan et al., 2023a, Table VI; Sec. VII) |
來源
Yuan et al., 2023a
(2023)SDV-LOAM: Semi-Direct Visual-LiDAR Odometry and MappingIEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9), pp. 11203-11220
DOI 10.1109/tpami.2023.3262817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
- 程式碼釋出:ZikangYuan/SDV-LOAM (GPL-2.0) https://github.com/ZikangYuan/SDV-LOAM
程式碼:https://github.com/ZikangYuan/SDV-LOAM(授權:GPL-2.0 (GitHub license metadata))。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。