Adaptive VINS-GNSS Fusion with Intelligent Switching and Failure Detection for Robust Autonomous Vehicle Localization
Mahmoud Adham, Wu Chen, Mostafa Mahmoud, Ahmed Mansour, Yang Yang, Yaxin Li · 2025
The integration of Visual-Inertial Navigation Systems (VINS) with the Global Navigation Satellite System (GNSS) has emerged as a pivotal approach for achieving robust and high-precision localization for autonomous vehicles in complex environments. While VINS provides local consistency through visual and inertial cues, its reliability is compromised in dynamic and low-texture scenarios. Conversely, GNSS ensures global positioning but suffers from signal degradation in urban canyons and occluded areas. To bridge this gap, we propose an adaptive fusion framework that dynamically integrates VINS and GNSS, leveraging their complementary strengths. The system intelligently switches between three operational modes: VINS-only, GNSS-only, and VINS-GNSS fusion based on real-time confidence assessment, ensuring continuous and accurate state estimation. Additionally, a novel dynamic object removal strategy is introduced, combining semantic awareness with multi-level geometric constraints to eliminate spurious visual features while preserving critical static landmarks. The backend incorporates an adaptive multi-layer VI-GNSS optimization framework that mitigates long-term drift and detects subsystem failures, enhancing overall robustness. Extensive experiments in urban environments validate the effectiveness of our approach, demonstrating superior positioning accuracy, resilience to environmental variations, and improved trajectory consistency.