Sensitivity analysis as a general tool for model optimisation – examples for trajectory estimation

Abstract

This paper outlines the general characteristics of variance-based sensitivity analysis and their advantages with respect to other concepts of sensitivity analysis. The main benefit are qualitative and quantitative correct results independent of the model characteristics. The author focuses on kinematic positioning as required for car navigation, driver assistance systems or machine guidance. The paper compares two different Kalman filter approaches using variance analysis and variance-based sensitivity analysis. The approaches differ with respect to their measurement quantities (input), their state quantities (output), as well as their dynamic vehicle model. The sensitivity analysis shows that each model has its different advantages and input-output relations. Furthermore it is shown that the variance-based sensitivity analysis is well suited to detect the share of the influence of the input quantities on the output quantities, here the estimated positions. Even more important, changes in deterministic and stochastic models lead to obvious effects in the respective variances and sensitivity measures. This emphasises the possibility to optimise the filter models by use of the variance-based sensitivity analysis.