DSA-GAN: Driving Style Attention Generative Adversarial Network for Vehicle Trajectory Prediction

Abstract

One of the main issues that potentially cause faults in ego-vehicle trajectory prediction is various styles of drivers. To deal with this problem, we propose Driving Style Attention Generative Adversarial Network (DSA-GAN), which can generate the trajectory of ego-vehicle conditioned on the driving style. This system can be adopted in many vehicles because it only needs CAN-bus data to predict the trajectory. The proposed architecture involves two stages, Driving style recognition and Trajectory prediction. In the Driving style recognition, Recurrence Plot (RP) transforms sequential data into images and the converted images are processed into the driving styles by Convolutional Neural Network (CNN). In the Trajectory prediction part, Conditional Generative Adversarial Network (CGAN) generates the multi-modal realistic trajectories from the distribution and these trajectories are conditioned by the driving style. In this paper, we predict more realistic and accurate trajectories than conventional prediction methods, even if a driver's driving style is not categorized by our defined classes.