Visual Odometry (Nistér et al.)
本文提出並命名「視覺里程計」，只用影像即時估計單一相機或立體相機的運動。前端在每張影像偵測 Harris 角點，以正規化互相關在視差限制內比對並做雙向一致性檢查，再把匹配串成軌跡。單目版以五點法估計三視角相對方向、三角化後以三點法求位姿並估計尺度；立體版則直接三角化再以三點法求位姿並以左右影像共同評分，全部用搶先式 RANSAC 與迭代精修。定期重新三角化形成「防火牆」，阻止錯誤傳播。
本頁內容
Nistér et al. define visual odometry as real-time camera motion estimation from video alone, combining Harris-corner tracking with normalised cross-correlation, minimal-solver preemptive RANSAC (5-point for monocular, 3-point for stereo), iterative refinement and periodic re-triangulation firewalls.
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)、monocular camera |
|---|---|
| 原文測試平台 | vehicle (autonomous ground vehicle with two front stereo heads; quantitative runs)、aerial (single camera mounted obliquely on an aeroplane; qualitative, Fig. 1) |
| 狀態估計 | frame-to-frame visual odometry: monocular scheme with three-view 5-point relative orientation, triangulation, 3-point resection and scale estimation, all in preemptive RANSAC with iterative refinement; stereo scheme with triangulation followed by 3-point pose from the left image, scored and refined on reprojection errors in both images; Cauchy robust cost; periodic re-triangulation as a firewall against error propagation (Sec. 4) |
| 資料關聯 | Harris corners without absolute thresholds (up to 5000 features in 10 x 10 buckets); 11 x 11 normalised cross-correlation within a disparity limit (typically 10% of image size); mutual-consistency check; matches linked into tracks (Secs. 2, 3) |
| 時間表示 | discrete video frames (about 13 Hz processing on the vehicle) |
| 去畸變 | 不適用 |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | none persistent; locally triangulated sparse 3D points only |
| 先驗資訊 | calibrated cameras; known 28 cm stereo baseline gives metric scale (Secs. 4.2, 5.1) |
| 可輸出幾何 | 6-DoF camera trajectory in metric scale (stereo); obstacle maps when combined with a separate stereo obstacle-detection module (Fig. 8) |
| 計算需求 | all processing at video rates on a 1 GHz Pentium III class machine with MMX-optimised feature detection and matching; about 13 Hz on the vehicle because of concurrent tasks (Secs. 1, 2, 5.1) |
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） | high precision Inertial Navigation System (model not reported) | 參考或真值量測 | 未標示 | orientation always from the inertial sensors | (Nistér et al., 2004, Secs. 5, 5.1, 5.3) |
| GNSS 接收器 | DGPS in RT-2 mode (model not reported) | 參考或真值量測 | 未標示 | up to 2 cm relative accuracy; part of the integrated INS/DGPS vehicle navigation system | (Nistér et al., 2004, Secs. 5, 5.1) |
| 雙目相機 | pair of synchronized analog cameras (two stereo heads on the vehicle; model not reported) | 方法輸入 | 未標示 | 50 deg horizontal field of view, 720 x 240 image fields, tilted about 10 deg to the side, 28 cm baseline | (Nistér et al., 2004, Sec. 5.1; Fig. 4) |
| 載具平台 | autonomous ground vehicle (mobile robotic platform) | 方法輸入 | 未標示 | two stereo heads mounted in the front | (Nistér et al., 2004, Fig. 4; Sec. 5) |
| 運算硬體 | 1 GHz Pentium III class machine | 執行運算平台 | 未標示 | all processing at video rates | (Nistér et al., 2004, Sec. 1) |
作者報告的優勢與限制
優勢
- Distance errors of 1.07%, 4.86% and 1.63% relative to DGPS over 186 to 366 m ground-vehicle runs (Table 1)
- Three tight laps totalling 184 m returned with only 4.1 m endpoint error (Fig. 5 caption)
- Frame-to-frame heading discrepancy to INS with standard deviation 0.50 to 0.59 deg and near-zero mean, suggesting unbiased yaw (Table 2)
- The stereo scheme works even without camera motion and gives metric scale from the baseline (Sec. 4.2)
- Trajectories stayed upright without any imposed motion prior (Sec. 5)
限制
- Dead reckoning: error accumulates over time like the INS heading (Sec. 5.3)
- Monocular frame selection for relative orientation is simplistic; a model-selection version was not yet fast enough (Sec. 4.1)
- Absolute orientation between stereo 3D point sets performed much worse because of depth uncertainty (Sec. 4.2)
- Scoring and optimising on one stereo image only leads to drift or torsion from calibration bias (Sec. 4.2)
- Quantitative evaluation limited to three outdoor trail runs of a few hundred metres (Table 1)
營建工程相關證據
論文未在營建環境測試；定量實驗為自主地面車在林地步道上數百公尺的三段行駛（分別名為 Loops、Meadow、Woods），並以 DGPS 與 INS 為參考。作為視覺里程計的奠基文獻，它說明以相機追蹤特徵、以最小解 RANSAC 估計位姿的基本流程，是後續工地視覺 SLAM 與相機點雲拼接的方法源頭；本身不產生地圖點雲（推論）。
原文驗證環境：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
Nistér et al., 2004
（2004）Visual odometryProceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2004), vol. 1, pp. 652-659 (Crossref; IEEE Xplore metadata lists the pages only as 'I', and the PDF carries no printed page numbers)
同儕審查已出版已讀全文經典查證後修正