SpeedNet: dual-channel CNN-GRU network with adaptive noise fusion for robust vehicle dead reckoning in GNSS-denied environments

Abstract To address error accumulation and dynamic disturbances in inertial navigation systems for ground vehicles operating in GNSS-denied environments, this paper proposes a collaborative navigation method integrating deep learning with adaptive filtering. First, we design a dual-channel CNN-gated recurrent unit speed estimation network (SpeedNet) that decouples inertial measurement unit (IMU) time-series data into independent 2D feature maps for acceleration and angular velocity channels. This extracts spatiotemporal correlations of translational and rotational motions while extending the traditional non-holonomic constraint (NHC) into 3D velocity constraints (NHC3) to establish vehicle kinematic priors. Second, a multimodal constraint dynamic adaptation mechanism is developed to autonomously adjust zero-velocity updates (ZUPT), zero angular-rate updates (ZIHR), and NHC3 filtering processes based on real-time vehicle motion states, resolving model mismatch caused by constraint failures in dynamic scenarios. Furthermore, a joint Bayesian dynamic estimation framework quantifies the impact of time-varying IMU noise characteristics on SpeedNet prediction errors and the unscented Kalman filter observation model in GNSS-denied environments, enabling cooperative optimization of process noise covariance matrix Q and measurement noise covariance matrix R . Experiments conducted on the KITTI dataset demonstrate that the proposed method achieves a drastic reduction in navigation error, outperforming the traditional method by 66.2% ± 3.2% in absolute trajectory error ( p < 0.001). This result underscores the fundamental limitation of conventional methods in GNSS-denied environments and highlights the transformative performance gain enabled by our deep-learning-enhanced, jointly-adaptive framework.