Lane change trajectory prediction considering driving style uncertainty for autonomous vehicles

Abstract

Lane change trajectory prediction is crucial for autonomous vehicles (AVs) to assess their own driving safety in advance. However, there are significant uncertainties in the implementation of such prediction, including different behaviors caused by agent–agent interaction and driving styles of drivers. While prototype trajectories can serve as a means to represent typical motion patterns and enhance trajectory prediction performance, their utilization in modeling motion patterns tends to overlook the influence of agent–agent interactions and vehicle dynamics. This paper proposes a fusion algorithm that considers driving style and vehicle dynamics to address these uncertainties. The algorithm involves a long short-term memory (LSTM) lane change behavior recognition model that mines key features of agent–agent interaction through the attention mechanism. The Gaussian process (GP) motion modeling trajectory prediction (GPMM-TP) algorithm considers the driving style of the prototype trajectory based on the behavior recognition results. To further improve short-term and long-term prediction, the interactive multi-model (IMM) algorithm is used to assign probability weights to the GP model and the Extended Kalman Filter (EKF) model based on prediction accuracy, taking into account the driving styles and the vehicle dynamics. The proposed algorithm provides a promising approach to improving the accuracy of lane change trajectory prediction for AVs, and its effectiveness is demonstrated using the HighD dataset.