Registration-based deskewing
作者以點對面 ICP 配準相鄰兩幀 LiDAR 點雲取得幀間運動，假設單幀掃描期間轉換參數的變化率固定、雷射發射間隔固定，依各點的發射順序線性內插出部分轉換，把每個點轉回該幀起始位姿，因此不需 IMU，也不需每點的實際時間戳記。作者另提出以參考平面為基準的去畸變位移指標 PRDT，處理點到平面距離在運動方向與目標表面平行時無法反映畸變的問題；並把同一模型反向用於動態 LiDAR 模擬器，以雷射發射頻率更新感測器位姿，產生帶運動畸變的合成點雲。室內驗證使用 UGV 搭載 Velodyne HDL-32E，並以 TLS 點雲作為參考。
本頁內容
IMU-free deskewing: the inter-frame motion from point-to-plane ICP between consecutive scans is linearly interpolated over the laser firing order to move each point back to the frame-start pose; adds the plane-referenced deskewing translation (PRDT) metric and a dynamic LiDAR simulator built by reversing the model; validated indoors with a UGV-mounted Velodyne HDL-32E against a TLS reference and in a synthetic street model.
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 | 3D LiDAR Velodyne HDL-32E (UoM indoor data; up to 72,000 points per cloud, 1187 clouds)、No IMU used by the method、Evaluation reference: TLS point cloud (scanner model not reported); the dataset also includes a BIM of the same area, but no evaluation against the BIM is reported |
|---|---|
| 原文測試平台 | UGV, indoor (third floor of the University of Melbourne engineering building; average 0.33 m/s and 14.85 deg/s)、Simulated vehicle at 40 to 100 km/h in a synthetic street 3D model (LiDAR at 10 fps) |
| 狀態估計 | Pairwise point-to-plane ICP between consecutive clouds gives the inter-frame transformation, followed by per-point linear interpolation of its rotation and translation parameters; no filter, smoother or pose-graph step |
| 資料關聯 | Point-to-plane ICP correspondences between consecutive point clouds (Low 2004); the authors state that any registration method could be substituted |
| 時間表示 | Constant rate of change of the transformation parameters within one frame and a constant interval between consecutive laser firings; each point's fraction of the frame motion comes from its firing-order index, so actual timestamps are not required |
| 去畸變 | IMU-free, registration-based: the ICP transformation between consecutive frames is linearly interpolated per point by firing order and each point is transformed back to the frame-start pose; the reverse process is used by the simulator to create distorted clouds |
| 迴圈閉合 | 不適用 |
| 全域最佳化 | none |
| 地圖表示 | 不適用 |
| 先驗資訊 | none |
| 可輸出幾何 | Deskewed point clouds; simulated skewed point clouds together with matching ideally deskewed clouds that serve as exact point-to-point ground truth |
| 計算需求 | For clouds of about 70,000 points: average registration 0.074 s, coordinate update below 0.001 s, total below 0.1 s per frame, which the authors state allows real-time use with a 10 Hz LiDAR; hardware not 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 | Velodyne HDL-32E | 資料集感測器 | UoM indoor data | up to 72,000 points per cloud; average 67,573 points over 1187 clouds; mounted on a moving UGV | (Zhao et al., 2024a, Sec. 4 Data (UoM indoor data)) |
| 地面雷射掃描儀（TLS） | TLS (Terrestrial Laser Scanner), model not reported | 參考或真值量測 | UoM indoor data | described as highly precise; 10 RANSAC planes manually selected from the TLS cloud for plane fitting | (Zhao et al., 2024a, Sec. 3 Evaluation metrics, Sec. 4 Data, Fig. 4 to 5) |
| 載具平台 | UGV (Unmanned Ground Vehicle), model not reported | 資料集感測器 | UoM indoor data | average linear speed 0.33 m/s, average turn rate 14.85 deg/s; frame 63 captured at 0.33 m/s and 10.23 deg/s | (Zhao et al., 2024a, Sec. 4 Data (UoM indoor data)) |
作者報告的優勢與限制
優勢
- No IMU and no per-point timestamps required; any registration method can supply the inter-frame motion (Sec. 1, Sec. 3)
- UoM indoor data: nearly 50% of point pairs moved more than 3 cm closer to the reference surface (PRDT) and about 5% by 10 cm or more (Sec. 4, Fig. 8)
- Frame 63: deskewed cloud about 5 cm closer to the TLS reference at larger cut-off thresholds, and the visible tear is reduced (Sec. 4, Fig. 9 to 10)
- About 0.074 s registration plus below 0.001 s coordinate update per ~70,000-point cloud, total below 0.1 s (Sec. 5)
- Simulator updates the sensor pose at the laser firing rate and yields skewed and ideally deskewed clouds with exact correspondence; simulated distortion agrees with the theoretical model (Sec. 3, Sec. 4, Fig. 7 and 11)
- Source code released (Sec. 1)
限制
- Outdoor conclusions rest on simulation and indoor tests; no real outdoor dataset with a matching TLS reference was available (Sec. 5)
- Relies on successful registration of consecutive clouds, which is difficult where distinctive static geometry is lacking in all three dimensions (Sec. 5)
- Moving objects may disturb registration and hence deskewing (Sec. 5)
- Indoor platform moved slowly (about 1 km/h), so indoor improvements are around 3 cm (Sec. 5)
- Assumes a constant rate of change of motion parameters within a scan and a constant firing interval (Sec. 3)
- No comparison with IMU-based deskewing; stated as future work (Sec. 6)
- PRDT discards point pairs not fitted to the same plane and excludes angles too close to 90 deg (threshold 10 deg used empirically) (Sec. 3, Sec. 4)
營建工程相關證據
未在施工現場測試；室內驗證位於大學工程館樓層，以 TLS 點雲作為參考（資料集另附同一區域的 BIM，但文中未見以 BIM 進行評估）。作者指出室內應用的精度要求較嚴，公分級改善也有意義；對缺少同步 IMU 的建築室內移動掃描，可作為去畸變與評估方法的參考（推論）。
原文驗證環境：已完工建築、獨立參考量測、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 3 個比較組，合計 8 筆紀錄。
Zhao et al., 2024a · Text Sec.5 Discussion 本方法 5 筆
資料集與序列UoM indoor data · all clouds
表格設定（擷取紀錄原文）：Summary statements in the Discussion on indoor improvement, simulated outdoor improvement and runtime for about 70,000-point clouds (Zhao et al., 2024a, Text Sec.5 Discussion)
PRDT values around 3 cm (indoor improvement)，UoM indoor data · all clouds
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 Zhao et al., 2024a 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 |
|---|---|---|
| proposed point cloud deskewing method本方法原文提出 | 3 cm有附註註記（擷取紀錄）：other: approximate ('around') | (Zhao et al., 2024a, Sec. 5 Discussion) |
Zhao et al., 2024a · Text Sec.4 indoor evaluation 本方法 2 筆
資料集與序列UoM indoor data · all 1187 clouds
表格設定（擷取紀錄原文）：PRDT computed for every skewed and deskewed point pair in the UoM indoor dataset with angle threshold 10 deg; percentages stated in text describing the cumulative distribution in Fig. 8 (Zhao et al., 2024a, Text Sec.4 indoor evaluation)
share of point pairs moved more than 3 cm closer to the reference surface (PRDT above 3 cm), stated as 'nearly 50%'，UoM indoor data · all 1187 clouds
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 Zhao et al., 2024a 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 |
|---|---|---|
| proposed registration-based deskewing (point-to-plane ICP)本方法原文提出 | 50%有附註註記（擷取紀錄）：other: approximate value ('nearly') stated in text | (Zhao et al., 2024a, Sec. 4 (Evaluation of deskewing on indoor dataset), Fig. 8) |
Zhao et al., 2024a · Text Sec.4 frame 63 本方法 1 筆
指標reduction of mean point-to-point distance to TLS at larger cut-off thresholds (deskewed 'about 5 cm closer')
資料集與序列UoM indoor data · frame 63 (0.33 m/s, 10.23 deg/s)
表格設定（擷取紀錄原文）：Mean point-to-point distance to the TLS reference cloud against cut-off threshold for frame 63; difference between skewed and deskewed clouds at larger cut-off thresholds stated in text (Zhao et al., 2024a, Text Sec.4 frame 63)
reduction of mean point-to-point distance to TLS at larger cut-off thresholds (deskewed 'about 5 cm closer')，UoM indoor data · frame 63 (0.33 m/s, 10.23 deg/s)
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 Zhao et al., 2024a 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 |
|---|---|---|
| proposed registration-based deskewing (point-to-plane ICP)本方法原文提出 | 5 cm有附註註記（擷取紀錄）：other: approximate improvement relative to the skewed cloud, not an absolute distance | (Zhao et al., 2024a, Sec. 4 (Evaluation of deskewing on indoor dataset), Fig. 10) |
來源
Zhao et al., 2024a
（2024）Registration‐based point cloud deskewing and dynamic lidar simulationThe Photogrammetric Record, 39(188), pp. 831-844
同儕審查已出版已讀全文近十年