Recursive Bayesian Estimation Applied to Autonomous Vehicles

This thesis presents an implementation of a sequential extended Kalman filter applied to position, velocity and attitude estimation of autonomous vehicles. The filter is self-tuning by the introduction of a particle swarm otimization which tunes the process noise covariance. The sensor fusion is adaptive through the means of corrector signals. It accepts correctors in position, velocity and attitude, in all possible combinations. The algorithm also includes an extended Kalman filter for quaternion update in order to make the estimations more robust when implementing sensor fusion flexibility. The filter architecture developed in this thesis is called the adaptive self-tuning extended Kalman filter, or the ASTEK filter. The algorithm was first tested in MATLAB/Simulink and then implemented and finalized in C++ in order to facilitate real-time performance. From testings on a truck, the RMS error for estimating position using a GPS corrector lies in the interval [10−4, 0.002] m, for velocity in [10−4, 0.02] m/s, and for the estimated attitude in [10−3, 0.21] degrees, depending on the road and driving mode. When using a 2D map corrector, that is correcting for x, y, and yaw, the RMS estimations of roll and pitch are higher and lying in the interval [1.1, 3.1]. However, it is kept stable as a result from the quaternion EKF, whereas the z-direction diverges as expected. The results show that the algorithm is able to produce estimations of high accuracy and that the corrector signals may vary dynamically. Moreover, the results show how different roads and driving modes influence the estimation and error evolution.