I. Introduction
Advanced driver-assistance systems have gained wide application in mass-production vehicles to improve driving safety and comfort. Among these intelligent systems, lateral driving-assistance systems (LDASs) provide a punctual warning (in lane departure warning system, LDW) or intervention (in lane keeping assist system, LKA) to prevent unintended lane departures and have great potential to avoid traffic accidents [1], [2]. A hierarchical structure is usually employed in the design of LDASs, which mainly consists of an upper level controller and a lower level controller, as presented in [2] and [3]. The upper layer anticipates vehicle trajectory and determines whether a warning or intervention is needed. The lower layer is used to regulate the vehicle states or alert human drivers with light, sound, or vibration. The research of this article lays an emphasis on the solution of vehicle trajectory prediction, an important basis in LDASs. However, because of the complex interaction between human drivers and vehicle dynamics, it is difficult to achieve an accurate long-term trajectory prediction. To avoid human–machine conflict and reduce the rate of false alarms and interventions, LDAS must achieve a comprehensive understanding of the driver’s characteristics and accurately predict vehicle states [4].