Trajectory Prediction for Human-driven Vehicles Based on a Combined Method of Deep Neural Network and Driving Risk Map

Abstract

Accurately predicting the trajectories of human-driven vehicles is critical for the safe and efficient driving of autonomous vehicles in mixed traffic environment. However, such prediction is challenging due to the vehicle motion uncertainty caused by the driver's intention and diverse driving styles. Moreover, the future movement of the vehicle is affected by other surrounding vehicles, and is constrained by the traffic lanes and road edges. To address the above problems, we propose a combined method of deep neural network and driving risk map for trajectory prediction of the human-driven vehicle (HV), which contains a conditional variational autoencoder (CVAE) to generate possible trajectories considering the uncertainty of vehicle motion and an interaction module to calculate the probability of the future trajectory of the vehicle in terms of impacts of the above traffic elements on the vehicle movement. Different from the method that only considers the influence among vehicle motions, a unified and explicit interaction model is established based on the driving risk map. Experiments using the published HighD dataset reveal that our method performs well in terms of the probability of the best predicted trajectory and prediction accuracy.