Adaptive Visual-Inertial Fusion for GNSS-Denied Environments Using Factor Graph Optimization

The performance of navigation systems would be highly degraded in GNSS-denied environments. In this work, an adaptive visual-inertial fusion method based on factor graph optimization is proposed to enhance the integrated navigation system. It presents a sliding-window framework that integrates IMU preintegration, visual odometry, and a novel MLP-based position selection factor, which dynamically evaluates the reliability of visual and inertial information to enable intelligent sensor fusion. Experimental results demonstrate that the proposed method achieves RMSE of 0.50 m and maximum positioning error of 1.31m, reducing RMSE by 52.8% and maximum error by 63.5% compared to EKF fusion, while outperforming fixed-weight FGO by an additional 20.6% in RMSE reduction.