Leveraging drivers' driving preferences into vehicle speed prediction using oriented hidden semi-markov model

Accurate vehicle speed prediction has important practical value to enhance fuel economy, drivability, and safety of intelligent vehicles. Current research on vehicle speed prediction mainly focuses on adapting to the dynamics, random and complex driving environment, while rarely takes drivers' driving preferences into account. In this paper, a learning-based prediction model consisted of an oriented Hidden Semi-Markov model (Oriented-HSMM) and an optimal preference speed prediction algorithm is proposed to leverage drivers' driving preferences into vehicle speed prediction. The Oriented-HSMM is developed to learn the spatial-temporal coherence of drivers' driving preference states under different traffic conditions and infer its long-term sequences in position domain. Based on these preference states, the optimal speed prediction algorithm using preference dynamics features is designed to retrieve the speed trajectory with maximal likelihood. To show its effectiveness, the proposed method is tested with the Next Generation Simulation (NGSIM) data on the US101 dataset comprising with the Hidden Markov model (HMM) and HSMM without considering driving preferences. Experiment results indicate that the proposed algorithm obtains the best performance with the mean absolute error (MAE) of 4.15 km/h and the root mean square error (RMSE) of 0.7603 km/h at 200 m prediction horizon.