DeepUKF-VIN: Adaptively-tuned Deep Unscented Kalman Filter for 3D Visual-Inertial Navigation based on IMU-Vision-Net

2025-01-01

Expert systems with applications

AI Summary

To address insufficient state estimation accuracy in three-dimensional visual-inertial navigation (VIN) for unmanned aerial vehicles under GPS-denied conditions, this paper proposes an end-to-end differentiable deep unscented Kalman filter (UKF) framework. The method tightly fuses monocular image sequences with inertial measurement unit (IMU) data and introduces, for the first time, an adaptive parameter-tuning deep UKF architecture. By jointly learning noise statistics and motion priors through an IMU-Vision-Net, it enables online filter parameter optimization and dynamic error covariance calibration. Evaluated on the EuRoC dataset (sequences MH_01–03), the proposed approach reduces absolute trajectory error (ATE) by 37% compared to conventional UKF, MSCKF, and VINS-Mono, while achieving a real-time inference speed of 42 FPS—demonstrating significant improvements in both estimation accuracy and computational efficiency.

Technology Category

Intelligent Robots: State EstimationMachine Learning: Calibration & Uncertainty QuantificationComputer Vision: Learning & Optimization for CV

Application Category

User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendationSearch and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingGraph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Problem

• 3D Space
• GPS-denied Environment
• UAV Navigation

Innovation

• DeepUKF-VIN
• Quaternion-based Filtering
• GPS-denied Environment Navigation