| Abstract | Robotic navigation system designers prefer optimization-based state estimators for applications that prioritize accurate state estimation over computational efficiency. Unlike single-estimator design, utilizing multiplemodel (MM) assumptions of the platform's dynamics has been shown to achieve robust performance in a dynamic environment. In MM optimization-based state estimators, performing state interaction within a computationally efficient framework is crucial for achieving real-time performance. This paper proposes a multiple-model horizon scenario tree (MM-HST) approach with model interaction for optimization-based state estimators. The problem is addressed by optimizing a sliding window (SW) of pre-integrated measurement history to determine the optimal state estimate considering two or more possible models. The different combinations in which the model can switch within a sliding window are maintained in a tree of solutions, where each solution within the window is denoted as a scenario. The MM scenario that results in minimum measurement residuals advances the SW optimizer by sliding forward the estimation window and marginalizing the out-of-window states. To evaluate the proposed method, a benchmark multi-model estimation target tracking problem involving constant velocity, constant acceleration, and constant turn rate models is utilized. Numerical simulations validate the performance of the proposed method in terms of accuracy, consistency, and sensitivity to trajectory dynamics while considering variations in its main design parameters. Compared to the state-of-the-art interacting multiple model (IMM) filtering approach, the MM-HST demonstrates improved estimation performance for changing dynamic scenarios. Experimental validation is conducted using a multiple-mode target tracking scheme performed in a motion capture room. The results indicate that the MM-HST with model interaction outperforms the stand-alone estimators used in the MM bank by interacting between the model scenarios in the horizon tree framework based on optimizing the minimum measurement residual. |
|---|
