Deep Neural Network-enhanced Integrated Navigation Using Global Navigation Satellite System and Inertial Navigation System for Vehicle Safety Testing

Abstract Reliable and accurate vehicle positioning is essential for intelligent transportation and active safety testing. However, in complex urban environments, global navigation satellite system (GNSS) signals are frequently degraded due to blockage and multipath effects, resulting in significant positioning errors. To address this issue, this paper proposes an adaptive GNSS/INS integration framework based on an Attention-gated recurrent unit (GRU) network. The method leverages inertial navigation system (INS) data to generate pseudo-GNSS observations during GNSS outages or under multipath interference, and integrates them within a Kalman filtering framework for robust state estimation. Extensive experiments are conducted in both simulation and real-world scenarios. In simulation, the proposed method significantly outperforms traditional Kalman filtering, reducing the positioning root mean square error (RMSE) from 37.234 m to 2.672 m, demonstrating strong robustness under GNSS-denied conditions. On the public UrbanNav dataset, the proposed approach achieves an RMSE as low as 0.953 m under outage conditions and 1.271 m under multipath interference, consistently outperforming state-of-the-art deep learning and filtering-based methods. Furthermore, real-world vehicle experiments show that the method reduces positioning errors by >70% compared with the traditional Kalman filter during GNSS degradation scenarios. These results demonstrate that the proposed Attention-GRU framework effectively mitigates INS drift and suppresses multipath-induced biases, providing a robust and accurate solution for GNSS/INS integrated navigation in challenging environments.