Adaptive vehicle longitudinal trajectory prediction for automated highway driving
Abstract
This paper describes an adaptive vehicle longitudinal trajectory prediction method for automated highway driving applications. A major strength of this method is that it can cope with highly dynamic situations in which the constant acceleration (CA) assumption cannot guarantee long term prediction accuracy. In this method, a quintic polynomial is used to model the longitudinal dynamics of a vehicle that is maneuvering. The decision to switch to it from the CA model is formulated as a maneuver detection problem. A maneuver is detected through monitoring measurement innovations of a Kalman filter that tracks target longitudinal states. The longitudinal jerk, as a dynamic characteristic of a maneuver is also estimated from measurement innovations. Finally the estimated jerk and context information are incorporated into the quintic polynomial model. The overall approach was tested on recorded human driving data from a simulator in a dynamic highway merging scenario. The results show the proposed method has higher prediction accuracy than the CA based method in such a dynamic scenario