A CGAN-based Model for Human-like Driving Decision Making

Abstract

Autonomous vehicles have attracted increasing interest in recent years and decision making is a critical task for autonomous driving. To improve both the efficiency and safety of autonomous driving in various traffic scenarios, the target vehicle needs to timely make rational driving decisions according to its historical driving state and environment state. Therefore, a driving model based on Conditional Generative Adversarial Networks (CGAN) is proposed to imitate human driving behavior, which can generate the mapping from the current state to the driving decision. The proposed model not only exploits the spatiotemporal relationship of the observable driving states by leveraging the Convolutional LSTM (ConvLSTM) network, but also learns the different weight values of hidden features by utilizing the self-attention mechanism. Simulation results indicate that the proposed model is capable of generating human-like driving decisions and outperforms other reference models based on real traffic data.