Automatic Tuning of the Unscented Kalman Filter and the Blind Tricyclist Problem: An Optimization Problem
Leonardo Azevedo Scárdua, José Jaime da Cruz · IEEE Control Systems · 2016
The extended Kalman filter (EKF) [1] has been a widely used nonlinear estimation tool for more than four decades [2], but it may perform poorly when the dynamic system is not almost linear on the time scale of the update intervals [3]?[5]. In such cases, the unscented Kalman filter (UKF) [6] has the potential to achieve better estimation performance than the EKF, while having computational complexity of the same order of magnitude [6]. To work properly, the UKF requires values for the three parameters of the (scaled) unscented transform (UT) [7]. The user is thus faced with a difficult task, for which there is little theoretical guidance. This fact has given rise to numerous heuristics for tuning of the UT parameters (see "Related Work" for a brief review of the relevant literature).