VKFus:Variational Inference as a Unifying Framework for Learned Absolute and Relative Pose Fusion

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Communication Engineering > Electronic Thesis & Dissertation > Item 987654321/98143

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98143

Title:	VKFus:Variational Inference as a Unifying Framework for Learned Absolute and Relative Pose Fusion
Authors:	陳咿汝;Chen, Yi-Ru
Contributors:	通訊工程學系
Keywords:	姿態估計;深度學習;擴展卡爾曼濾波;變分貝葉斯推斷;Pose Estimation;Deep Learning;Extended Kalman Filter;Variational Bayesian Inference
Date:	2025-08-16
Issue Date:	2025-10-17 12:24:49 (UTC+8)
Publisher:	國立中央大學
Abstract:	本論文提出 VKFus，一個透過變分推斷來統一學習式絕對與相對姿態融合的彈性框架，旨在應對具挑戰性環境下的定位問題。我們的方法透過證明擴展卡爾曼濾波器 (Extended Kalman Filter, EKF) 優化與證據下界 (Evidence Lower Bound, ELBO) 最大化在高斯假設下的等價性，建立了一個嚴謹的數學基礎。為感測器融合提供了理論依據，使本方法有別於經驗性方法。 VKFus 框架能夠適應多樣的感測器配置，展現了其靈活性。該系統採用一個絕對姿態迴歸 (Absolute Pose Regression, APR) 分支以實現無漂移的全域定位，以及一個相對姿態迴歸 (Relative Pose Regression, RPR) 分支來處理包含慣性測量單元 (Inertial Measurement Unit, IMU) 量測數據在內的多模態感測器資料，以進行運動估計。兩個分支均整合了基於注意力的不確定性估計模組，能同時預測相機姿態與其不確定性，從而根據環境條件對不同模態進行自適應加權。實驗結果顯示，在具挑戰性的 EuRoC 資料集上，VKFus 達到了 0.57 公尺的平均均方根誤差 (Root Mean Square Error, RMSE)，相較於近期基於 EKF 的方法，性能提升高達 41%。在 KITTI 資料集上，我們的方法在定位精度上展現了 40% 的提升。這些一致的性能提升，驗證了 VKFus 作為一個可靠導航系統解決方案的有效性。;This paper presents VKFus, a flexible framework that unifies learned absolute and relative pose fusion through variational inference for positioning under challenging conditions. Our approach establishes a mathematically rigorous foundation by demonstrating the equivalence between Extended Kalman Filter (EKF) optimization and Evidence Lower Bound (ELBO) maximization under Gaussian assumptions. This formulation provides a theoretical basis for sensor fusion, distinguishing the approach from empirical methods. The VKFus framework demonstrates flexibility by accommodating diverse sensor configurations. The system employs an Absolute Pose Regression (APR) branch for drift-free global positioning and a Relative Pose Regression (RPR) branch that processes multi-modal sensor data including IMU measurements for motion estimates. Both branches incorporate attention-based uncertainty estimation modules that simultaneously predict camera poses and uncertainties, enabling adaptive weighting between modalities based on environmental conditions. Experimental results demonstrate state-of-the-art performance. On the challenging EuRoC dataset, VKFus achieves an average RMSE of 0.57m, an improvement of up to 41% over recent EKF-based methods. On the KITTI dataset, our approach demonstrates a 40% improvement in positioning accuracy. The consistent performance improvements validate VKFus′s effectiveness as a reliable solution for navigation systems.
Appears in Collections:	[Graduate Institute of Communication Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	3	View/Open

社群 sharing

Loading...