Traj-LO

LiDAR-only odometry with a piecewise-linear continuous-time trajectory optimized in a sliding window, coupling point-to-plane geometry with smoothness constraints so no separate deskewing is needed.

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。

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。

作者報告的優勢與限制

優勢

• On Hilti 2021 handheld sequences continuous-time methods outperform discrete-time LO with the non-repetitive Livox MID70 even without IMU, but MID70 accuracy is lower than with the Ouster OS0-64 and Traj-LO with MID70 alone diverged on the Lab sequence (Sec. IV-D; Table III)
• Adding the vertical LiDAR (L1+L2) removes the large z-direction errors on the tnp sequences (ATE 0.049, 0.040, 0.049 m versus 0.505, 0.607, 0.101 m with L1 only) and on spms lowers ATE and avoids the spms_02 divergence (Sec. IV-C; Fig. 4; Table II)
• KITTI online benchmark 0.58% translation and 0.0014 deg/m rotation error (Sec. IV-B)
• Maps aggressive motion beyond the IMU range on the Point-LIO rotating platform where FAST-LIO fails (Sec. IV-E; Fig. 5)

限制

• Single horizontal LiDAR shows larger ATE during jerky take-off and landing phases (arXiv v1 Sec. IV-C and Fig. 4 on spms_03; this paragraph is not retained in the version of record)
• Constant velocity within segments may not hold for fast motion, requiring short segments (Sec. III-B) (author-stated assumption)
• Assumes the odometry starts from a stationary state and uses the first 0.3 s of points to initialize the map (Sec. III-E)
• Single-LiDAR Traj-LO diverged on NTU VIRAL spms_02 and has ATE of 0.505 m and 0.607 m on tnp_01 and tnp_02 (Table II)
• Odometry only; a complete LiDAR-only SLAM is left for future work (Sec. V)

營建工程相關證據

作者描述其使用的 Hilti 2021 資料集包含辦公室、實驗室與施工環境之室內序列及施工現場之室外序列，地面真值來自 Hilti PLT 300 全測站或動作擷取系統。表 III 的 Cons2 依 Hilti 2021 官方序列名稱推定為 Construction Site Outdoor 2（推論），該序列官方地面真值為全測站量測的 3 自由度稀疏點位；Traj-LO 在 Cons2 的 ATE 為 0.065 m（OS0-64）、0.135 m（MID70）與 0.063 m（兩者併用）。僅有軌跡層級評估，未評估點雲幾何。

原文驗證環境：公開基準、施工中工地、獨立參考量測

來源

• Zheng & Zhu, 2024

  Xin Zheng, Jianke Zhu（2024）Traj-LO: In Defense of LiDAR-Only Odometry Using an Effective Continuous-Time TrajectoryIEEE Robotics and Automation Letters, 9(2):1961-1968

  DOI 10.1109/lra.2024.3352360arXiv 2309.13842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：Traj-LO (arXiv v1) https://arxiv.org/abs/2309.13842
• 程式碼釋出：kevin2431/Traj-LO https://github.com/kevin2431/Traj-LO

程式碼：https://github.com/kevin2431/Traj-LO（授權：MIT (LICENSE file checked)）。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。

回到方法圖鑑