Neural-Aided State Estimation for Nonlinear Vehicle Tracking Using KalmanNet and GRU.

Abstract

Accurate tracking of vehicles is a key problem in autonomous driving, especially when the maneuvers are complex and/or under high noise sensing measurements. Traditional model-based estimators, such as the EKF, can easily be affected by linearization errors and are dependent on precise mathematical modelling and carefully calibrated noise characteristics. In this paper, we propose an enhanced learning-aided state estimator called KalmanNet GRU that maintains the recursive property of Kalman filtering and introduces a neural dynamics learner based on a GRU (gated recurrent unit) cell and a learnable Kalman gain in order to achieve robustness against nonlinearities and measurement noises. The performance of the proposed approach is demonstrated via the HDVT problem with nonlinear range–bearing radar measurements at an inverse sensing noise level of 20 dB. The experimental results reveal that KalmanNet significantly outperforms the EKF as it can attain an average test MSE of −7.91 dB in contrast to 9.85 dB resulted by the EKF, which means achieving nearly 17.76 dB improvement in performance. The model proposed also has a stable estimation performance and yet holds an acceptable processing time for real-time autonomous navigation. These results verify the utility of recursive filtering by data-driven dynamics learning for nonlinear state estimation in ITS.