Adaptive Techniques for Tuning of Process Noise in the Kalman filter

Abstract

This thesis work focuses on tuning the process noise of the Kalman filter for sur rounding object detection and speed estimation in automotive driver assistance ap plications. We have created a simple road scenario in Matlab, where an ego-vehicle follows a target-car moving at different speeds and yaws in different road scenarios to collect the required data for the Kalman filter. The tracking algorithms help in predicting the future position of moving objects based on the measurements of RADAR detections. The RADAR sensors attached to the ego-vehicle directly mea sure the range, azimuth angle and range rate of the target object. We considered the compensated range rate and azimuth angle measurements from the RADAR sensors. Since the RADAR measurements are noisy, the Kalman filter is used to obtain a better result. It is well known that the covariance matrices of the process noise (Q) and the measurement noise (R) have a significant impact on the performance of the Kalman filter in estimating dynamic states. We begin by implementing the con ventional ad-hoc approaches to estimating the covariance matrices. However, these approaches for estimating the covariance matrices, may not be suitable for achieving the best filter performance. To address this problem, we propose a adaptive filtering approach to estimate Q based on innovation and residual to improve the accuracy of the dynamic state estimation of the extended Kalman filter (EKF). We also present theoretical investigation methods for optimizing the algorithm when it is to be im plemented in an embedded platform. Following that we present the results obtained in our work after comparing the fixed and the adaptive Kalman filter, which clearly points that the adaptive Kalman Q is better in its response time and adaptivity.