VKFPos: A Learning-Based Monocular Positioning with Variational Bayesian Extended Kalman Filter Integration

2025-01-31

AI Summary

To address the challenge of fusing absolute and relative pose uncertainties in monocular visual localization, this paper proposes a variational Bayesian inference-based end-to-end extended Kalman filter (VB-EKF) framework. Methodologically, it introduces VB-EKF as a deep architectural component for the first time, implementing an APR/RPR dual-branch network that jointly regresses poses and predicts covariance matrices; theoretical derivation yields a decoupled posterior structure enabling uncertainty-aware spatiotemporal localization. Key contributions include: (1) tight integration of variational Bayesian EKF with deep learning, and (2) explicit uncertainty modeling via dual-branch covariance prediction. Experiments demonstrate state-of-the-art accuracy in single-frame absolute pose estimation; for sequential localization, the method significantly outperforms pure APR baselines and conventional EKF across multiple indoor and outdoor datasets, achieving an average reduction of 23.6% in localization error.

Technology Category

Intelligent Robots: State EstimationReasoning under Uncertainty: Relational Probabilistic ModelsMachine Learning: Calibration & Uncertainty Quantification

Application Category

Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphsSearch and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingSecurity and Privacy: Browser and end-point security

Abstract

This paper addresses the challenges in learning-based monocular positioning by proposing VKFPos, a novel approach that integrates Absolute Pose Regression (APR) and Relative Pose Regression (RPR) via an Extended Kalman Filter (EKF) within a variational Bayesian inference framework. Our method shows that the essential posterior probability of the monocular positioning problem can be decomposed into APR and RPR components. This decomposition is embedded in the deep learning model by predicting covariances in both APR and RPR branches, allowing them to account for associated uncertainties. These covariances enhance the loss functions and facilitate EKF integration. Experimental evaluations on both indoor and outdoor datasets show that the single-shot APR branch achieves accuracy on par with state-of-the-art methods. Furthermore, for temporal positioning, where consecutive images allow for RPR and EKF integration, VKFPos outperforms temporal APR and model-based integration methods, achieving superior accuracy.

Problem

• Single Camera Localization
• Accuracy Improvement
• Uncertainty Prediction

Innovation

• VKFPos
• Combined Absolute and Relative Pose Regression
• Extended Kalman Filter for Uncertainty Prediction