Basalt
Basalt 採兩層架構整合視覺慣性里程計與全域一致建圖。下層為立體視覺 VIO，以 KLT 光流追蹤 FAST 角點，在滑動視窗中聯合最佳化重投影與 IMU 預積分誤差，並以首次估計 Jacobian 做部分邊際化。當關鍵影格被邊際化時，作者用非線性因子還原（NFR）把邊際化先驗近似成關鍵影格間的相對位姿因子與橫滾俯仰因子。上層以 ORB 特徵在關鍵影格間匹配並做光束法平差，結合這些因子得到重力對齊的全域地圖，不需估計每個關鍵影格的速度與偏差。
本頁內容
Basalt pairs a KLT-based stereo VIO fixed-lag smoother with keyframe bundle adjustment on ORB features, transferring VIO information to the global map through relative-pose and roll-pitch factors obtained by non-linear factor recovery, which yields gravity-aligned, globally consistent visual-inertial maps.
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、IMU |
|---|---|
| 原文測試平台 | UAV (EuRoC MAV machine hall and Vicon room sequences) |
| 狀態估計 | two layers: (1) stereo VIO as fixed-lag smoother (Gauss-Newton over 7 pose-only keyframes and 3 latest states with velocity and biases) combining reprojection and preintegrated IMU terms, Schur-complement partial marginalisation with first-estimate Jacobians; (2) visual-inertial mapping as keyframe bundle adjustment of ORB landmarks plus relative-pose and roll-pitch factors recovered from the VIO marginalisation prior by non-linear factor recovery (KL-divergence minimisation) (Secs. IV, V) |
| 資料關聯 | VIO: FAST corners in a 50 pixel grid (80 to 120 features) tracked by pyramidal inverse-compositional KLT with SE(2) patch warp and locally scaled SSD, forward-backward consistency check; mapping: ORB features detected and matched between keyframes (Secs. IV-A, V-A, VI-a) |
| 時間表示 | discrete frames with IMU preintegration between consecutive frames (Sec. IV-B3) |
| 去畸變 | 不適用 |
| 迴圈閉合 | implicit, through ORB keypoint matching between keyframes in the global bundle adjustment (Sec. V) |
| 全域最佳化 | keyframe bundle adjustment with recovered non-linear relative-pose and roll-pitch factors (yaw and absolute-position factors dropped); map is gravity aligned (Secs. V-A to V-C) |
| 地圖表示 | keyframe poses and ORB landmark positions (inverse distance with stereographic bearing parameterisation) in a gravity-aligned global map; VIO landmarks hosted in keyframes (Secs. IV-B, V-A) |
| 先驗資訊 | camera projection functions (intrinsics) and camera-IMU extrinsics are assumed static and known from calibration (Sec. IV-B); the paper does not state how IMU noise parameters are obtained |
| 可輸出幾何 | VIO pose for every frame; globally consistent, gravity-aligned keyframe trajectory and sparse ORB landmark map (Fig. 1) |
| 計算需求 | Intel E5-1620 (4 cores, 8 threads), highly parallel implementation; VIO 7.83 ms per frame on average (5.5 to 9.4 ms); mapping 52.8 ms per keyframe; MH_05 (114 s) processed in 19.2 s VIO plus 9.7 s mapping (Sec. VI-d; Table II) |
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 |
|---|---|---|---|---|---|
| 慣性量測單元（IMU） | EuRoC MAV IMU (model not reported in this paper) | 資料集感測器 | EuRoC MAV | 原文未報告 (IMU measurements preintegrated between consecutive frames) | (Usenko et al., 2020, Secs. IV-B3, VI) |
| 雙目相機 | EuRoC MAV stereo camera (model not reported in this paper) | 資料集感測器 | EuRoC MAV | MH_05 has 2273 stereo frames over 114 s; V2_03 has more than 400 missing frames for one camera | (Usenko et al., 2020, Sec. VI-d; Table I note) |
| 運算硬體 | Intel E5-1620 | 執行運算平台 | 未標示 | 4 cores, 8 virtual cores; implementation uses all available CPU resources | (Usenko et al., 2020, Sec. VI-d) |
作者報告的優勢與限制
優勢
- Proposed VIO best on eight of ten EuRoC sequences among VIO methods (Table I; Sec. VI-b)
- Proposed VI mapping reaches 0.02 to 0.10 m RMS ATE on EuRoC and clearly beats VI ORB-SLAM on machine-hall sequences with long keyframe intervals (Table I)
- Properly weighted recovered factors beat pure BA and identity-weighted factors in accuracy and robustness (Table I; Sec. VI-c)
- Global optimisation state is about 2.5 times smaller because velocities and biases are not estimated (Sec. VI-d)
- Fast: VIO 7.83 ms per frame and whole MH_05 about four times faster than real time (Sec. VI-d)
限制
- Camera intrinsics and camera-IMU extrinsics are assumed static and known from calibration (Sec. IV-B)
營建工程相關證據
論文未在營建工地測試；只在 EuRoC 室內 MAV 資料評估。其重力對齊的全域一致關鍵影格地圖，以及相對低的運算量，對以立體相機與 IMU 做施工現場巡檢定位有參考價值；輸出為稀疏 ORB 地標，若要得到工程點雲需額外稠密重建（推論）。
原文驗證環境：公開基準、獨立參考量測
報告的性能數據
性能數據仍在分批查證，目前尚未收錄此方法的報告值。
來源
Usenko et al., 2020
（2020）Visual-Inertial Mapping With Non-Linear Factor RecoveryIEEE Robotics and Automation Letters, 5(2):422-429
DOI 10.1109/lra.2019.2961227arXiv 1904.06504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：Visual-Inertial Mapping with Non-Linear Factor Recovery (arXiv v1 to v3) https://arxiv.org/abs/1904.06504
- 程式碼釋出：Basalt (GitLab VladyslavUsenko/basalt; project page vision.in.tum.de/research/vslam/basalt) https://gitlab.com/VladyslavUsenko/basalt
程式碼：https://gitlab.com/VladyslavUsenko/basalt（授權：BSD-3-Clause (GitLab project licence 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。