An Online Learning Based Extended Kalman Filtering Approach for Intelligent Vehicles Localization During Short-term GNSS Outages

Real-time and accurate localization is a prerequisite for intelligent vehicles control. GNSS is an important information source for localization. However, GNSS signal may be short-term blocked by large buildings and tunnels inevitably. Therefore, it is a practical issue to retain localization accuracy during short-term GNSS outages. By improving the modeling accuracy of the vehicle motion and sensor measurements, localization is expected to maintain a satisfactory performance during GNSS short-term outages. In this paper, dual neural extended kalman filtering approach (DN-EKF) is introduced to compensate for the unmodeled errors of vehicle motion and statistical modeling error of sensor measurement noise, and consequently improves estimator accuracy. Experiments on our test platform have demonstrated the effectiveness of proposed method during GNSS short-term outages. It is worth noting that the proposed method in this paper is open-ended. Therefore, it can be easily integrated with other solutions to further improve the performance of localization.