Neural network based extended Kalman filter for localization of mobile robots

This paper studies the localization of a mobile robot based on neural network based extended Kalman filter (NNEKF) algorithm. Extended Kalman filter (EKF) is used to fuse the information acquired from both the robot optical encoders and ultrasonic sensors in order to estimate the current robot position and orientation. Then the error covariance of the EKF is tracked by the covariance matching technique. When the output of the matching technique does not meet the certain condition, a neural network is employed to modify the system noise covariance matrix. Simulation results demonstrate that, with the comparison to the odometry and the standard EKF method under the same error divergence condition, the proposed algorithm effectively improves the accuracy of the localization of the mobile robot system and prevents the filter divergence.