VIL-SLAM
VIL-SLAM 把三個模組串接：緊耦合的雙目視覺慣性里程計以固定滯後的位姿圖平滑器估計運動，並以 IMU 頻率輸出位姿；LiDAR 建圖模組用這些位姿為每個點去畸變，再以 LOAM 式邊緣與平面特徵做掃描對地圖配準；迴圈閉合先以視覺詞袋偵測候選並用 EPnP 求初始約束，再以稀疏 LiDAR 特徵點的 ICP 精修，最後用 iSAM2 增量最佳化全域位姿圖，並把修正後的位姿即時回饋給建圖模組重新定位。作者以 Faro 地面雷射掃描為參考評估地圖，並在隧道與走廊等 LiDAR 退化場景顯示視覺慣性先驗的幫助。
本頁內容
Loosely chained stereo VIO (tightly coupled fixed-lag smoother), LOAM-style LiDAR mapping seeded and dewarped by IMU-rate VIO poses, and LiDAR-enhanced visual loop closure (BoW detection, EPnP, sparse-feature ICP) with incremental iSAM2 pose-graph optimization, targeting LiDAR-degenerate tunnels and hallways.
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 | stereo camera pair (two megapixel cameras; model not reported)、16 scan-line 3D LiDAR (model not reported)、IMU at 400 Hz (model not reported) |
|---|---|
| 原文測試平台 | 原文未報告 (custom sensor platform shown in Fig. 1(a); carrier platform not stated) |
| 狀態估計 | loosely coupled chain: tightly coupled stereo VIO as a fixed-lag smoother over the most recent N stereo frames (IMU pre-integration factors and structureless vision factors, Levenberg-Marquardt, Schur-complement marginalization, GTSAM) provides IMU-rate motion priors to LOAM-style LiDAR mapping; a global pose graph of LiDAR mapping poses with LiDAR odometry and loop constraint factors is optimized incrementally with iSAM2 (Secs. V-VII) |
| 資料關聯 | visual: KLT tracking of stereo matches, Shi-Tomasi corners with ORB descriptors and brute-force stereo matching; LiDAR: edge and planar feature points registered scan-to-map by point-to-line (two closest edge points) and point-to-plane (three closest surface points) distances as in LOAM (Secs. IV, VI-B) |
| 時間表示 | discrete stereo-frame states in the VIO; LiDAR points dewarped with IMU-rate VIO poses; custom microcontroller circuit synchronizes cameras, LiDAR, IMU and computer by simulating GPS time signals (Secs. V, VI-A, VIII-A) |
| 去畸變 | each LiDAR point dewarped to the end-of-scan time using the closest IMU-rate VIO poses (Sec. VI-A, Eq. 6) |
| 迴圈閉合 | yes; visual Bag-of-Words detection (DBoW3) of key images within a time threshold, descriptor matching to reject false positives, EPnP initial constraint, then ICP refinement on sparse LiDAR feature points of the key scans (LibPointMatcher) (Sec. VII-A, VII-B) |
| 全域最佳化 | incremental global pose-graph optimization with iSAM2 over all LiDAR mapping poses; corrected poses sent back in real time so LiDAR mapping re-localizes and updates its feature map (Sec. VII-C, VII-D) |
| 地圖表示 | sparse LiDAR feature map (all previous edge and surface feature points) for registration; dense map produced in post-processing by stitching dewarped scans with the best estimated poses, reported as 1 cm voxel dense maps near real time (Sec. III; abstract) |
| 先驗資訊 | none |
| 可輸出幾何 | loop-closure-corrected 6-DoF LiDAR poses in real time and a dense point-cloud map near real time (abstract; Fig. 5) |
| 計算需求 | 4 GHz computer with 4 physical cores on the platform; EuRoC VIO results obtained in real time on a desktop with a 3.60 GHz i7-4790 CPU (Secs. VIII-A, VIII-C); no per-module timing reported |
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 scan-line LiDAR (model not reported) | 方法輸入 | 未標示 | 3D LiDAR on the custom platform | (Shao et al., 2019, Sec. VIII-A) |
| 地面雷射掃描儀（TLS） | Faro time-of-flight laser scanner (model not reported) | 參考或真值量測 | 未標示 | scans used as the reference model for mean registration error | (Shao et al., 2019, Sec. I; Sec. VIII-B) |
| 慣性量測單元（IMU） | IMU (model not reported) | 方法輸入 | 未標示 | 400 Hz | (Shao et al., 2019, Sec. VIII-A) |
| 雙目相機 | two megapixel cameras (model not reported) | 方法輸入 | 未標示 | stereo pair on the custom platform | (Shao et al., 2019, Sec. VIII-A; Fig. 1(a)) |
| 運算硬體 | 4 GHz computer with 4 physical cores (model not reported) | 執行運算平台 | 未標示 | onboard computer of the custom platform | (Shao et al., 2019, Sec. VIII-A) |
| 運算硬體 | Intel i7-4790 desktop CPU | 執行運算平台 | EuRoC MAV | 3.60 GHz; EuRoC VIO results obtained in real time | (Shao et al., 2019, Sec. VIII-C) |
| 其他 | custom microcontroller-based time synchronization circuit | 方法輸入 | 未標示 | synchronizes cameras, LiDAR, IMU and computer by simulating GPS time signals | (Shao et al., 2019, Sec. VIII-A) |
作者報告的優勢與限制
優勢
- Succeeded in the hallway and tunnel tests, where LOAM accumulated large error or failed because of degeneracy along the traversal direction (Sec. VIII-B; Table I)
- Lower final drift error than LOAM on highbay and hallway and equal on outdoor; lower mean map registration error than LOAM against Faro scans on highbay, hallway and huge loop (Table I)
- Loop closure lowered the final drift to 0.05% in the hallway and 0.08% in the tunnel (Sec. VIII-B)
- LOAM failed the huge-loop test after re-entering the highbay, while VIL-SLAM reached 0.01% final drift without a loop closure being triggered (Sec. VIII-B)
- Stereo VIO succeeded on all EuRoC sequences with accuracy comparable to three state-of-the-art methods (Sec. VIII-C; Fig. 8)
限制
- In the tunnel, degeneracy still caused error along the traversal direction; a single loop constraint did not fully remove the doubled map (Sec. VIII-B; Fig. 7)
- Featureless hallway walls under-constrain the VIO and misalign the map until loop closure (Sec. VIII-B)
- VIO and LiDAR mapping are only loosely coupled, so LiDAR does not correct IMU biases (Sec. IX)
- Loop-closure ICP uses sparse features between scans; the authors expect scan-to-map matching would give better constraints (Sec. IX)
- Only compared with one LiDAR baseline (LOAM) on custom datasets; KITTI not used because the evaluation sequences lack inertial data (Sec. VIII)
營建工程相關證據
測試場景為倉庫式高挑空間、無特徵走廊、隧道與戶外道路，未在施工中工地；地圖精度以 Faro 飛時測距雷射掃描為參考，先對齊再計算地圖點到參考模型最近點的平均距離，屬於有獨立幾何參考的點雲評估，這種以 TLS 為真值的做法可直接借鏡到施工點雲驗收。走廊與隧道的 LiDAR 退化問題也常見於施工中的地下或長廊空間（推論）。
原文驗證環境：已完工建築、地下或隧道、獨立參考量測、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 2 個比較組，合計 12 筆紀錄。
Shao et al., 2019 · Table I 本方法 10 筆
表格設定（擷取紀錄原文）：Author-collected sequences that start and end at the same point; FDE = final drift error of LiDAR mapping odometry (no loop closure) as % of distance; MRE = mean distance from map points to the closest point of Faro reference scans after aligning the map; LOAM = laboshinl/loam_velodyne implementation; '-' not finished, 'x' missing data (Shao et al., 2019, Table I)
FDE final drift error，VIL-SLAM custom datasets · Highbay (total length 118, unit not printed)
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 Shao 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(Shao et al., 2019, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| VIL-SLAM本方法原文提出 | 0.08% | (Shao et al., 2019, Table I) |
| LOAM | 0.56% | (Shao et al., 2019, Table I) |
Shao et al., 2019 · Text Sec.VIII-B 本方法 2 筆
資料集與序列VIL-SLAM custom datasets · Hallway
表格設定（擷取紀錄原文）：Final drift error after loop closure, stated in the text (Shao et al., 2019, Text Sec.VIII-B)
FDE after loop closure (loop detected twice near the endpoint)，VIL-SLAM custom datasets · Hallway
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 Shao et al., 2019 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 |
|---|---|---|
| VIL-SLAM (with loop closure)本方法原文提出 | 0.05% | (Shao et al., 2019, Sec. VIII-B) |
來源
Shao et al., 2019
（2019）Stereo Visual Inertial LiDAR Simultaneous Localization and Mapping2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 370-377
DOI 10.1109/iros40897.2019.8968012arXiv 1902.10741
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：Stereo Visual Inertial LiDAR Simultaneous Localization and Mapping (arXiv v1, submitted version) https://arxiv.org/abs/1902.10741