Data-Driven Merging of Car-Following Models for Interaction-Aware Vehicle Speed Prediction

The paper presents a prediction framework applied to vehicle speed prediction in multi-lane traffic. In general, the framework combines components of (heuristic) model knowledge, probabilistic estimation and machine learning techniques in order to benefit from the different advantages of the respective methods. The idea of the approach is merging several models suitable for simple environments (called basis models) to a bigger model suitable for complex environments (called overall model). Since the number of basis models is limited due to faster execution time, system boundaries are determined, which implicates a selection of the relevant model inputs. For improved prediction performance, the values of the model parameters are estimated online via recursive Bayesian estimation. Moreover, data-driven models are integrated for adaptive weighting of the basis models in order to represent time-varying behavior over the prediction horizon. In the application of the framework to vehicle speed prediction, single-lane car-following models are used as basis models. The determination of the system boundaries is based on a decision tree. Model parameter estimation is realized with a particle filter implementation and the data-driven models for weighting the basis models are realized as support vector machine (SVM) regression models. Experimental results of the suggested vehicle speed prediction framework show an improved performance of the approach compared to a related baseline approach.