An efficient tuning framework for Kalman filter parameter optimization using design of experiments and genetic algorithms

摘要

The Extended Kalman Filter (EKF) is currently a dominant sensor fusion method for mobile devices, robotics, and autonomous vehicles. Its performance heavily depends on the selection of EKF parameters. Therefore, the optimal selection of parameters is a critical factor in EKF design and use. In this paper, a methodical and efficient method of EKF parameter tuning is presented. The tuning framework uses nominal parameters generated by Gauss Markov (GM) and Allan Variance (AV) methods that are tuned by Genetic Algorithms (GA) accelerated by Design of Experiments (DoE). This framework has been implemented in MATLAB and tested using simulations and real data under a tightly coupled EKF that fuses IMU and GNSS measurements of a self-driving car provided by the Blackberry QNX company. The results demonstrate that GA-tuned parameters increase accuracy substantially over nominally tuned parameters, and that the DoE technique consistently improves the convergence behavior of the GA.