Point-LIO
Point-LIO 在每一個 LiDAR 點或 IMU 取樣到達時，就以不迭代的流形擴展卡爾曼濾波（on-manifold EKF）進行傳播與更新，里程計輸出可達 4 至 8 kHz，並從架構上避免掃描內的運動畸變。作者把角速度與線加速度擴增為一階積分隨機過程狀態，將 IMU 量測視為系統輸出，飽和的 IMU 通道直接略過，因此旋轉超出 IMU 量程時仍能估計位姿。每個點以 ikd-Tree 中五個最近鄰擬合平面，計算一維點到平面殘差。
本頁內容
Point-LIO fuses every LiDAR point and IMU sample at its own timestamp in a non-iterated on-manifold EKF, treating IMU readings as outputs of a stochastic-process-augmented kinematic model so saturated channels can be skipped; one-dimensional point-to-plane residuals against an ikd-Tree map give 4 to 8 kHz odometry without frame-level deskewing.
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 | Livox Avia solid-state LiDAR with built-in BMI088 IMU (own experiments)、public benchmarks with Livox Horizon, Velodyne HDL-32E and VLP-16 plus their IMUs |
|---|---|
| 原文測試平台 | robot car (DJI RoboMaster 2019 AI)、step-motor rotating platform、pendulum、racing quadrotor UAV、self-rotating UAV、public vehicle and campus sequences (utbm, ulhk, liosam, lili) |
| 狀態估計 | Tightly coupled on-manifold EKF (IKFoM toolbox), deliberately not iterated; 24-dimensional state on SO(3) x R21 with attitude, position, velocity, gyroscope and accelerometer biases, gravity, and angular velocity and linear acceleration modelled as first-order integrator processes; each LiDAR point or IMU sample is propagated and fused at its own timestamp, and saturated IMU channels are skipped |
| 資料關聯 | Each point is projected with the propagated pose; its five nearest map points within 5 m in the ikd-Tree are fitted to a plane; if any neighbour lies more than 0.1 m from the plane the point is added to the map without an update, otherwise a one-dimensional point-to-plane residual updates the state |
| 時間表示 | Discrete-time state propagated and updated at every measurement time (each LiDAR point or IMU sample), with the kinematic model discretized over each inter-measurement interval; because drivers deliver packaged scans, points and IMU samples in a package are sorted by timestamp and processed one by one |
| 去畸變 | No explicit deskewing step: each point is fused at its own sampling time so frame-level motion distortion does not arise; wall-thickness views show thinner walls than FAST-LIO2 (qualitative, Figs. 5 to 7) |
| 迴圈閉合 | 原文未報告 |
| 全域最佳化 | 原文未報告 |
| 地圖表示 | ikd-Tree incremental k-d tree point map from FAST-LIO2; local map size 2000 m, spatial downsampling 0.25 m, rebalancing thresholds 0.6 and 0.5, parallel rebuild threshold 1500 points; public benchmarks use FAST-LIO2 default mapping parameters with 1:4 temporal downsampling of raw points |
| 先驗資訊 | none |
| 可輸出幾何 | Point map accumulated in the ikd-Tree (points inserted at the updated pose) and odometry at 4 to 8 kHz; no export format or map accuracy evaluation reported |
| 計算需求 | DJI Manifold 2-C7 (Intel i7-8550U 1.8 GHz, 8 GB RAM): average 20.27 ms per scan on 12 public sequences versus 20.19 ms for FAST-LIO2 (Table 7), about 9 microseconds per point; nearest-neighbour search is sequential while FAST-LIO2 uses four threads; Khadas VIM3 Pro ARM board (Cortex-A73 2.2 GHz, 4 GB) onboard the self-rotating UAV: 14.63 ms per 50 Hz package |
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 | Livox Avia | 方法輸入 | 未標示 | solid-state, 70.4 deg x 77.2 deg circular FoV, non-repetitive scanning, 230,000 points/s, packages at 10 to 100 Hz | (He et al., 2023a, Sec. 5.2; Fig. 3a) |
| LiDAR | Livox Horizon | 資料集感測器 | lili (LILI-OM dataset) | solid-state 3D LiDAR | (He et al., 2023a, Sec. 6) |
| LiDAR | Velodyne HDL-32E | 資料集感測器 | utbm and ulhk | spinning LiDAR | (He et al., 2023a, Sec. 6) |
| LiDAR | VLP-16歸入：Velodyne VLP-16 | 資料集感測器 | liosam (LIO-SAM dataset) | spinning LiDAR | (He et al., 2023a, Sec. 6) |
| 慣性量測單元（IMU） | BMI088 (built-in IMU of Livox Avia) | 方法輸入 | 未標示 | 200 Hz; measuring range 35 rad/s and about 30 m/s2 (Sec. 5.5.1); range set to 17.5 rad/s on the self-rotating UAV (Sec. 7.2) | (He et al., 2023a, Sec. 5.2; Sec. 5.5.1; Sec. 7.2) |
| 相機 | first-person-view (FPV) camera | 資料集感測器 | Point-LIO own sequences (Livox Avia sensor suite) | aligned with LiDAR FoV; used only for visual illustration | (He et al., 2023a, Sec. 5.2; Fig. 3a; Sec. 7.1) |
| 載具平台 | RoboMaster 2019 AI robot car | 方法輸入 | Point-LIO own sequences (Livox Avia sensor suite) | sensor suite mounted on chassis without vibration absorber | (He et al., 2023a, Sec. 5.2; Sec. 5.3; Fig. 3b) |
| 載具平台 | rotating platform driven by Nimotion STM4260A step motor | 方法輸入 | Point-LIO own sequences (Livox Avia sensor suite) | peak yaw rate 75 rad/s in Satu-1 | (He et al., 2023a, Sec. 5.2; Sec. 5.5.1; Fig. 3c) |
| 載具平台 | pendulum (sensor suite on a rope) | 方法輸入 | Point-LIO own sequences (Livox Avia sensor suite) | circling motion, acceleration up to 40 m/s2 | (He et al., 2023a, Sec. 5.2; Sec. 5.5.2; Fig. 3d) |
| 載具平台 | racing quadrotor drone | 方法輸入 | 未標示 | thrust-to-weight ratio up to 5.4; carries Livox Avia and FPV camera; angular velocity up to 59.37 rad/s | (He et al., 2023a, Sec. 7.1; Fig. 18a) |
| 載具平台 | self-rotating single-actuated UAV | 方法輸入 | 未標示 | average yaw rate about 25 rad/s; Livox Avia facing front | (He et al., 2023a, Sec. 7.2; Fig. 18b) |
| 運算硬體 | DJI Manifold 2-C7 | 執行運算平台 | 未標示 | 1.8 GHz quad-core Intel i7-8550U, 8 GB RAM | (He et al., 2023a, Sec. 5.6; Sec. 6) |
| 運算硬體 | Khadas VIM3 Pro | 執行運算平台 | 未標示 | 2.2 GHz quad-core Cortex-A73, 4 GB RAM; runs Point-LIO onboard in real time | (He et al., 2023a, Sec. 7.2) |
| 其他 | Vicon Tracker (five Vicon markers on the sensor suite) | 參考或真值量測 | Point-LIO own sequences (Livox Avia sensor suite) | recorded at 300 Hz for bandwidth analysis | (He et al., 2023a, Sec. 5.2; Sec. 5.4) |
作者報告的優勢與限制
優勢
- 4 to 8 kHz odometry (6955 Hz average on Odo) and bandwidth above 150 Hz versus 100 Hz for FAST-LIO2 (Table 2)
- Survives IMU saturation after the initial stage: Satu-1 rotation and translation RMSE 4.60 deg and 0.233 m, where FAST-LIO2 and Point-LIO-input diverge once the IMU saturates; Satu-2 4.42 deg and 0.0990 m, with consistently lower errors than FAST-LIO2 and Point-LIO-input (Sec. 5.5)
- Park drift 0.080 m versus 1.242 m for FAST-LIO2 (Table 1)
- Best RMSE on 4 of 5 and lowest drift on 5 of 7 public sequences with one parameter set (Tables 5 and 6)
限制
- README requires LiDAR-IMU synchronization and per-point timestamps, and IMU saturation values must be configured (README, Important notes A-C)
- Fails when the motion already exceeds the IMU range at start (37.68 rad/s start failed, Table 3); rotation RMSE grows with initial angular velocity
- Worse than FAST-LIO2 and LILI-OM on the long lili_8 sequence, attributed to untuned shared parameters (Sec. 6.1.2)
- Point-wise processing is emulated after sorting packaged scans because drivers do not stream single points (Sec. 5.1)
- Odometry only: no loop closure or global map optimization module is described; Sec. 8 presents Point-LIO as an odometry (inference from system scope)
營建工程相關證據
未在施工現場或施工資料集上測試；實驗涵蓋機器人車於公園、廣場與走廊、實驗室旋轉台與擺錘、無人機，以及 utbm、ulhk、liosam、lili 公開序列，只提供軌跡、漂移與執行時間層級的證據，未評估點雲幾何精度。
原文驗證環境：受控實驗、公開基準
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 12 個比較組，合計 61 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 8 組列在最後，並連到性能比較頁。
Liu et al., 2026 · Table 2 (odometry without LC) 本方法 13 筆
指標absolute trajectory error (RMSE, centimeters)
表格設定（擷取紀錄原文）：Hilti handheld sequences (Hesai XT-32, BMI085 400 Hz); ATE exported from the Hilti evaluation website; odometry without loop closure; all methods with default parameters (Liu et al., 2026, Table 2 (odometry without LC))
absolute trajectory error (RMSE, centimeters)，Hilti handheld sequence exp01-construction (name per Table C1) · hilti01
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Liu et al., 2026 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Liu et al., 2026, Table 2 (odometry without LC))
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LeGO-LOAM | 9.1 cm | (Liu et al., 2026, Table 2) |
| LiLi-OM | 6.2 cm | (Liu et al., 2026, Table 2) |
| LINS | 6.5 cm | (Liu et al., 2026, Table 2) |
| LIO-SAM | 7.4 cm | (Liu et al., 2026, Table 2) |
| FAST-LIO2 | 1.3 cm | (Liu et al., 2026, Table 2) |
| Faster-LIO | 1.1 cm | (Liu et al., 2026, Table 2) |
| Point-LIO本方法 | 1.1 cm | (Liu et al., 2026, Table 2) |
| Our (Odom)原文提出 | 1.3 cm | (Liu et al., 2026, Table 2) |
| Our (Odom+LM)原文提出 | 0.8 cm | (Liu et al., 2026, Table 2) |
He et al., 2023a · Table 3 本方法 7 筆
指標RMSE of rotation (deg)
表格設定（擷取紀錄原文）：Spinning experiment started at different initial yaw rates; IMU range 35 rad/s; rotation RMSE against Vicon (He et al., 2023a, Table 3)
RMSE of rotation (deg)，Point-LIO own sequences (Livox Avia sensor suite) · start 6.28 rad/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 He et al., 2023a 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 |
|---|---|---|
| Point-LIO本方法原文提出 | 6.9 deg | (He et al., 2023a, Table 3) |
He et al., 2023a · Table 6 本方法 7 筆
指標drift (meters)
表格設定（擷取紀錄原文）：End-to-end drift (m) on sequences that start and end at the same place; lili uses Livox Horizon, ulhk Velodyne HDL-32E, liosam VLP-16; LILI-OM tuned per lili sequence (He et al., 2023a, Table 6)
drift (meters)，lili · lili_6
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
- 僅報告範圍
- 未報告（沒有數值，不是 0）
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 He et al., 2023a 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(He et al., 2023a, Table 6)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Point-LIO本方法原文提出 | 無數值僅報告範圍註記（擷取紀錄）：reported as <0.1 | (He et al., 2023a, Table 6) |
| FAST-LIO2 | 無數值僅報告範圍註記（擷取紀錄）：reported as <0.1 | (He et al., 2023a, Table 6) |
| LILI-OM | 0.8 m | (He et al., 2023a, Table 6) |
| LIO-SAM | 無數值未報告註記（擷取紀錄）：not reported in table (dash) | (He et al., 2023a, Table 6) |
| LINS | 無數值未報告註記（擷取紀錄）：not reported in table (dash) | (He et al., 2023a, Table 6) |
He et al., 2023a · Table 4 本方法 6 筆
指標time consumption per scan (ms)
表格設定（擷取紀錄原文）：Average total time per scan (ms); Park, Square, Corridor at 10 Hz, Odo, Satu-1, Satu-2 at 100 Hz; dash marks sequences where the LIO fails (He et al., 2023a, Table 4)
time consumption per scan (ms)，Point-LIO own sequences (Livox Avia sensor suite) · Park
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 He et al., 2023a 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(He et al., 2023a, Table 4)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| FAST-LIO2硬體：DJI Manifold 2-C7 (Intel i7-8550U 1.8 GHz quad-core, 8 GB RAM) | 39.84 ms | (He et al., 2023a, Table 4) |
| Point-LIO-input (ablation)硬體：DJI Manifold 2-C7 (Intel i7-8550U 1.8 GHz quad-core, 8 GB RAM) | 32.09 ms | (He et al., 2023a, Table 4) |
| Point-LIO本方法原文提出硬體：DJI Manifold 2-C7 (Intel i7-8550U 1.8 GHz quad-core, 8 GB RAM) | 32.94 ms | (He et al., 2023a, Table 4) |
其他比較組
來源
He et al., 2023a
（2023）Point‐LIO: Robust High‐Bandwidth Light Detection and Ranging Inertial OdometryAdvanced Intelligent Systems, 5(7):2200459
同儕審查已出版已讀全文近十年
相關版本
- 程式碼釋出：hku-mars/Point-LIO https://github.com/hku-mars/Point-LIO
程式碼：https://github.com/hku-mars/Point-LIO（授權：LICENSE file contains LOAM/Livox-derived BSD-3-Clause-style text (checked); overall project licensing not further verified）。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。