EKF-DT LocBump: Enhanced Real-Time 3D Localization and Terrain Bump Prediction with Sensor Fusion

In today's vehicle localization based on GNSS, IMU, and LiDAR sensor fusion, the extended Kalman filter (EKF) is widely used as the optimal generalized solution and much work has been directed towards improving its performance through various methods. However, the prediction of the state and extent of vehicle bumps in the height domain has rarely been addressed. When the vehicle bumps, i.e., the noise distribution, deviate significantly from the Gaussian distribution, the performance of the EKF localization is extremely limited: it is in a completely strongly nonlinear Gaussian state space (SS) system. In view of this, this paper investigates the combination of eleven machine learning (ML) algorithms with EKF, and finally proposes the EKF-DT system, a hybrid combination of a well-tuned decision tree (DT) algorithm and the classical EKF, aiming at correcting the performance of EKF by ML in order to realize accurate, real-time prediction of three-dimensional coordinates, in particular, high-precision trace (coordinate) prediction of the target in the height domain. EKF-DT solves the problem of traditional localization in sensor fusion. DT solves the legacy of traditional EKF in the sensor fusion field that cannot realize high-precision trace prediction of targets over the height domain, and further improves the EKF performance in trace prediction of longitude and latitude. In addition, EKF-DT has been proved to be excellent for this high-noise task in real-time and performance-compatible evaluations.