Early Intention Prediction of Lane-Changing Based on Dual Gaussian-Mixed Hidden Markov Models

Adjacent lane-changing is one of the most dangerous maneuvers which may lead to rear-end crash, uncomfortable braking and sharp steering. If the autonomous driving system can predict the potential latent lane-changing intentions of surrounding vehicles in advance, the driver will have more time to make reasonable response. In this paper, we focus on how to give accurate and reliable prediction for latent lane-changings, especially before the vehicles merge into the target lanes. A prediction model based on dual Gaussian-mixed hidden Markov models is developed to exploit the advantages of different features more effectively. Since there is no comprehensive criteria to evaluate the accuracy and predictability performance simultaneously, we propose two new metrics for quantitative analysis as supplement to the classical indicators. Comparative validation on Next Generation Simulation (NGSIM) database shows that our model has a high recognition accuracy of 93.05% for lane-changing intention with earlier prediction over the existing homologous methods.