Deep Learning-Enhanced Adaptive Statistical Nonholonomic Constraint Navigation Method for GNSS-Denied Environments

GNSS signal outages severely degrade vehicle integrated navigation accuracy, with positioning errors reaching tens of meters in environments such as urban tunnels. Traditional non-holonomic constraint (NHC) methods face difficulties adapting to complex motion dynamics, while odometry systems add cost and complexity. This paper proposes DL-ASNHC, a deep learning-enhanced adaptive statistical non-holonomic constraint navigation method. The approach designs a lightweight deep learning model, RCTNet, achieving high-precision forward velocity prediction in the vehicle frame (v-frame) with RMSE of 0.057 m/s. Motion state classification determines lateral velocity constraint parameters, which combine with vertical velocity constraints to form a complete pseudo-3D velocity observation model. An adaptive covariance mechanism based on multi-level anomaly detection (MAD, chi-square test, relative change rate) and Huber robust estimation dynamically optimizes filter parameters to handle complex motion variations. During 120s GNSS outage tests, DL-ASNHC achieves a minimum horizontal positioning RMSE of 0.12 m, representing 93.8% improvement over traditional schemes and 74.1% average improvement over other deep learning methods; the 90th percentile of positioning errors is 0.32 m. In tested scenarios, this method enables effective virtual odometry as an alternative to physical sensors, providing a cost-effective solution for GNSS-denied navigation.