Tag-based VIO for indoor construction UAVs

On-manifold (SE(3)) EKF that fuses onboard UAV velocity odometry with pixel measurements of AprilTag corners whose poses are known in the BIM frame, giving drift-free global localization for an off-the-shelf Parrot Bebop 2; validated in a Vicon lab and a BIM-enabled photo-realistic simulation.

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。

Tag-based VIO for indoor construction UAVs 的技術屬性
感測輸入	forward-looking monocular camera of the Parrot Bebop 2 (rectified 856 x 480 at about 30 Hz)、onboard IMU and odometry velocities of the Bebop 2 (about 5 Hz, upsampled to the image rate)
原文測試平台	UAV (Parrot Bebop 2, off-the-shelf, no hardware modification)、simulation (Parrot-Sphinx with Gazebo)
狀態估計	On-manifold extended Kalman filter on SE(3) with left perturbation; prediction from onboard translational and rotational velocities, correction from pixel coordinates of the four corners of each detected AprilTag whose pose is known in the BIM frame (Sec. 4)
資料關聯	AprilTag detection and ID decoding (AprilRobotics implementation) gives explicit tag-corner correspondences; no natural-feature matching
時間表示	discrete time steps at the image rate
去畸變	不適用 (camera and IMU only)
迴圈閉合	none (global tag measurements bound drift)
全域最佳化	none
地圖表示	no map is built; tag poses in the BIM reference frame serve as landmarks
先驗資訊	AprilTag family, size, IDs and global poses known a priori in the BIM coordinate system; camera intrinsics and camera-vehicle extrinsics from calibration
可輸出幾何	6-DoF UAV pose and covariance in the BIM frame; no point cloud
計算需求	runs on a ground station in ROS; hardware and timing 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。

原文使用的設備
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）	Bebop2 onboard IMU (odometry velocities)	方法輸入	未標示	onboard odometry about 5 Hz, boosted to the image rate	(Kayhani et al., 2022, Sec. 5.1, 5.1.1)
相機	Bebop2 forward-looking camera	方法輸入	未標示	rectified 856 x 480 images at about 30 Hz; focal length about 520 px	(Kayhani et al., 2022, Sec. 5.1.1, Sec. 6.2)
載具平台	Parrot Bebop2	方法輸入	未標示	compact off-the-shelf UAV with onboard flight controller, IMU, sonar and vertical camera for height, forward-looking camera; no hardware modification	(Kayhani et al., 2022, Sec. 5.1)
其他	Vicon motion capture system	參考或真值量測	未標示	sub-millimetre accuracy, above 200 Hz; used as ground truth and for closed-loop control	(Kayhani et al., 2022, Sec. 5.2.1)
其他	AprilTag 36h11 tags (6)	方法輸入	未標示	0.165 m x 0.165 m, letter-size paper, global pose in BIM frame known a priori	(Kayhani et al., 2022, Table 3)
其他	Parrot-Sphinx simulator with Gazebo	方法輸入	未標示	photo-realistic BIM-enabled simulation with simulated IMU, ultrasound, vertical and front cameras	(Kayhani et al., 2022, Sec. 5.1.2, Sec. 5.2.2)

作者報告的優勢與限制

優勢

Position RMSE as low as 2 to 5 cm in laboratory and simulation experiments (Abstract; Sec. 8)
Handles tag-blind zones by relying on odometry prediction and recovers after the next tag detection while remaining consistent within 3-sigma bounds (Sec. 6.1)
Tag-corner measurements and the on-manifold formulation gave more accurate and smoother estimates than an Euler-angle EKF using direct tag poses (Sec. 7.2, Fig. 19)
Requires only a camera and IMU on a compact commercial UAV and no mapping session (Sec. 1, 7)

限制

Tag measurements become unreliable beyond a distance threshold (under 4 m for 16.5 cm tags, 856 x 480 images, focal length about 520 px) (Sec. 6.2)
Manual tag placement and replacement is tedious and subject to installation errors that are not modelled (Sec. 7.2, 8)
Tags may be occluded or damaged on site; paper tags may curl (Sec. 2.3, 7.1.1)
Estimates in tag-blind zones depend on odometry quality and drift quickly with IMU-only prediction (Sec. 7.1.5)
Not yet validated on an actual construction site; state estimates were not used in the control loop (Sec. 6, 8)

營建工程相關證據

方法以施工中室內低紋理、反覆變動的環境為動機，並假設標籤位置已登錄在 BIM 中；實際驗證只在 Vicon 實驗室與依 BIM 產生的施工場景模擬中進行，作者把真實工地驗證列為未來工作（Sec. 5、Sec. 8）。

原文驗證環境：模擬、受控實驗、獨立參考量測

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 2 個比較組，合計 4 筆紀錄。

Kayhani et al., 2022 · Text Sec.6.1 本方法 2 筆

指標RMSE of 3D position estimates

資料集與序列

表格設定（擷取紀錄原文）：Simulation, planar trajectory with a tag-blind zone (experiment 3), BIM-enabled Parrot-Sphinx and Gazebo environment, six 0.165 m 36h11 AprilTags, camera-to-tag distance 1.5 to 4.5 m (Kayhani et al., 2022, Text Sec.6.1)

RMSE of 3D position estimates，BIM-enabled simulation (Parrot-Sphinx + Gazebo) · Planar (exp. 3), including take-off and landing

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 Kayhani et al., 2022 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。

統計量：均方根誤差（RMSE）；對齊方式：原文未報告；單位：m；場景：simulated indoor construction environment

數值與出處
方法（原文寫法）	報告值	出處
proposed tag-based on-manifold EKF本方法原文提出	0.0198 m	(Kayhani et al., 2022, Sec. 6.1)

Kayhani et al., 2022 · Text Sec.6.2 本方法 2 筆

指標RMSE in position

資料集與序列

表格設定（擷取紀錄原文）：Laboratory, 3D circular trajectory of radius 1 m (experiment 5), Vicon ground truth, camera-to-tag distance 2.2 to 4.2 m; value labelled in the text as 'including' take-off and landing disruptions (Kayhani et al., 2022, Text Sec.6.2)

RMSE in position，laboratory flight arena with Vicon · 3D circular (exp. 5), as labelled 'including take-off and landing'

統計量：均方根誤差（RMSE）；對齊方式：原文未報告；單位：m；場景：indoor laboratory

數值與出處
方法（原文寫法）	報告值	出處
proposed tag-based on-manifold EKF本方法原文提出	0.0348 m有附註註記（擷取紀錄）：other: the text pairs 0.0348 m with 'including' and 0.2602 m with 'excluding' disruptions, which looks reversed; values kept as written	(Kayhani et al., 2022, Sec. 6.2)

來源

Kayhani et al., 2022
Navid Kayhani, Wenda Zhao, Brenda McCabe, Angela P. Schoellig（2022）Tag-based visual-inertial localization of unmanned aerial vehicles in indoor construction environments using an on-manifold extended Kalman filterAutomation in Construction, 135:104112
DOI 10.1016/j.autcon.2021.104112
同儕審查已出版已讀全文近十年

回到方法圖鑑