Uncertainty-Aware Deep Kinematics and ESKF Fusion for Robust Vehicle Dead Reckoning

Abstract

Achieving reliable vehicle localization in GNSS outage environments is a critical challenge. Traditional Dead Reckoning (DR) systems using low-cost Micro-Electro-Mechanical Systems (MEMS) Inertial Measurement Units (IMUs) inevitably suffer from rapid error accumulation. This paper proposes an uncertainty-aware deep kinematics and Error-State Kalman Filter (ESKF) fusion framework. A hybrid one-dimensional convolutional neural network and gated recurrent unit (1D-CNN-GRU) model, trained with a Gaussian Negative Log-Likelihood (GNLL) loss, predicts body-frame velocity and data-dependent variance from raw IMU data. In the ESKF update, the learned forward velocity is used as the neural longitudinal measurement, while the lateral channel is constrained by the vehicle non-holonomic assumption. The learned forward variance is mapped to the measurement noise covariance to adaptively regulate the Kalman gain. Evaluated on a 394.34 m KITTI GNSS outage stress case with complex cornering, the proposed method suppresses position drift. In this stress case, the method reduces the terminal error by 77.39% relative to the ESKF + NHC baseline and yields a terminal drift rate of 7.32%. These results indicate that, although the learned velocity uncertainty is not fully calibrated in this stress case, it can still serve as a relative measurement-quality cue for vehicle dead reckoning under the evaluated GNSS outage condition.