The Extended Kalman Filter (EKF) is known to be an effective solution for the Simultaneous Localization and Mapping (SLAM) problem. However, the accuracy and the performance of the EKF-based SLAM are highly dependent on the knowledge of the noise statistics of the observation and process models. Specifically, wrong knowledge of the additive Gaussian noise covariance matrix in the observation model can dramatically degrade the quality and even the stability of the filtering process.
In this paper, a novel method based on Genetic Algorithms (GA) is proposed to tune the covariance parameters of the observation model. The innovation sequences which are the differences of the sensor measurements and the corresponding predicted estimations at different times are used to calculate the actual innovation covariance matrix. Then the actual innovation covariance is compared with the mathematical innovation covariance which is based on the statistics of the noise and is used in the update step of the Kalman filtering. If the mismatch between the theoretical covariance and the actual covariance is reduced, the statistics of the noise are more accurate. The proposed online adaptation solution utilizes the GA optimization to find the best covariance values those that minimize the mismatch.
The effectiveness of the proposed solution is demonstrated on a simulated mobile robot by solving the feature-based SLAM problem with initially untuned parameters of the observation statistics. The comparisons show that the GA-based adaptation EKF can successfully improve the performance of the EKF-based SLAM.
https://www.ece.unb.ca/petersen/pubs/SaPeLi11/