StereoScan (LIBVISO2)
StereoScan 以立體影像在單一 CPU 上即時建立三維地圖。前端以斑點與角點遮罩偵測特徵，用 Sobel 響應的稀疏 SAD 比對，並要求左右影像與前後影格四張影像形成環狀匹配；再以 RANSAC 包覆的高斯牛頓法最小化左右影像重投影誤差估計自我運動，並用等加速度卡爾曼濾波平滑速度。另一執行緒以 ELAS 計算稠密視差，依估計位姿把三維點重投影到目前影格，貪婪地合併並平均重複點，產生一致的點雲。
本頁內容
StereoScan (LIBVISO2) combines fast circular sparse feature matching and RANSAC reprojection-error stereo odometry at 25 fps with ELAS dense stereo at 3 to 4 fps and a greedy reprojection-based point fusion to build consistent 3D point clouds in real time on a CPU.
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 (calibrated, rectified) |
|---|---|
| 原文測試平台 | mobile platform recording the Karlsruhe stereo sequences (inner-city scenes、sequence names '2009_09_08_drive_00xx'、the vehicle is not described in the paper) |
| 狀態估計 | frame-to-frame stereo egomotion: Gauss-Newton minimisation of left and right reprojection errors of triangulated features inside RANSAC (50 iterations of 3-point samples), refinement on all inliers, then a constant-acceleration Kalman filter on the velocity (Sec. III-B) |
| 資料關聯 | blob and corner features from 5 x 5 masks with non-maximum and non-minimum suppression; SAD of quantised Sobel responses at 16 sparse locations of an 11 x 11 window; circular matching over left and right images of two frames with 1 pixel epipolar tolerance; Delaunay-neighbourhood support filtering; two-pass search narrowed per 50 x 50 pixel bin; bucketing to 200 to 500 features (Secs. III-A, III-B) |
| 時間表示 | discrete stereo frames (10 fps in the Karlsruhe data) |
| 去畸變 | 不適用 |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | point-based 3D model: ELAS disparity maps converted to 3D and greedily fused by reprojecting previous points into the current image and averaging points that fall on valid disparities (Sec. III-D) |
| 先驗資訊 | calibrated stereo rig with rectified images (Sec. III) |
| 可輸出幾何 | visual odometry trajectory at 25 fps and a fused dense 3D point cloud updated from new depth maps at 3 to 4 fps (abstract; Sec. III) |
| 計算需求 | two CPU threads: feature matching 36.6 ms plus visual odometry 4.3 ms per frame (about 25 fps) in the online setting; ELAS dense stereo 3 to 4 fps on a single i7 core at 3.0 GHz for about 0.5 MP images; appending one disparity map to the model usually under 50 ms (Fig. 6; Secs. III-C, III-D) |
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 |
|---|---|---|---|---|---|
| GNSS 接收器 | OXTS RT 3003 GPS/IMU | 參考或真值量測 | Karlsruhe dataset (cvlibs.net) | 'weak' ground truth; errors up to two metres possible in inner-city scenarios | (Geiger et al., 2011, Sec. IV-B) |
| 雙目相機 | stereo camera of the Karlsruhe dataset (model not reported) | 資料集感測器 | Karlsruhe dataset (cvlibs.net) | 1344 x 391 pixels, 10 fps, calibrated and rectified | (Geiger et al., 2011, Sec. IV) |
| 運算硬體 | i7 CPU | 執行運算平台 | 未標示 | single core at 3.0 GHz used for ELAS dense stereo; the pipeline assumes two CPU cores | (Geiger et al., 2011, Secs. III, III-C) |
作者報告的優勢與限制
優勢
- Visual odometry takes 4.3 ms versus about one second for Kitt et al. (2010) on 200 matches, a speed-up of more than 200 (Sec. IV-B; Fig. 6b)
- Multi-stage matching lowers running time while increasing the number of matches (Sec. IV-A)
- Localisation accuracy comparable to Kitt et al. on the Karlsruhe sequences despite much faster runtime (Sec. IV-B; Fig. 7, plotted only)
- First system reported by the authors to process about one-megapixel stereo images online on a single CPU (Sec. I)
- Greedy fusion reduces stored points and averages measurement noise over frames (Sec. III-D)
限制
- No loop closure or global optimisation; sequences are short and the soft reset of outdated points is not handled (Sec. III-D)
- Reconstruction targets static scenes; handling dynamic objects is future work (Secs. III, V)
- GPS/IMU reference is weak ground truth with errors up to about two metres in inner-city areas, so accuracy is judged only qualitatively in plots (Sec. IV-B)
- Three-dimensional reconstruction is evaluated only qualitatively (Sec. IV-C)
營建工程相關證據
論文未在營建工地測試；資料為 Karlsruhe 市區戶外立體影像序列，以 OXTS GPS/IMU 作弱真值，重建品質只做定性展示。它說明只用相機也能即時產生具公制尺度的融合點雲，對低成本工地巡檢或車載掃描有參考價值，但缺乏迴圈閉合與點雲精度量化，不能直接視為工程量測（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
Geiger et al., 2011
（2011）StereoScan: Dense 3d reconstruction in real-time2011 IEEE Intelligent Vehicles Symposium (IV), pp. 963-968
DOI 10.1109/ivs.2011.5940405程式碼
同儕審查已出版已讀全文經典查證後修正
相關版本
- 程式碼釋出：LIBVISO2 (C++ library with MATLAB wrappers; the cvlibs.net page gives this IV 2011 paper as the LIBVISO2 citation); dense stereo from LIBELAS https://www.cvlibs.net/software/libviso/
- 前身方法：Visual odometry based on stereo image sequences with RANSAC-based outlier rejection scheme (Kitt, Geiger, Lategahn, IV 2010; the freely available visual odometry library used as the baseline, cited as [16]) 10.1109/IVS.2010.5548123
程式碼：https://www.cvlibs.net/software/libviso/（授權：GNU GPL (LIBVISO page: code published under the GNU General Public License; version not stated)）。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。