MEDIRL: Predicting the Visual Attention of Drivers via Maximum Entropy Deep Inverse Reinforcement Learning

Inspired by human visual attention, we introduce a Maximum Entropy Deep Inverse Reinforcement Learning (MEDIRL) framework for modeling the visual attention allocation of drivers in imminent rear-end collisions. MEDIRL is composed of visual, driving, and attention modules. Given a front-view driving video and corresponding eye fixations from humans, the visual and driving modules extract generic and driving-specific visual features, respectively. Finally, the attention module learns the intrinsic task-sensitive reward functions induced by eye fixation policies recorded from attentive drivers. MEDIRL uses the learned policies to predict visual attention allocation of drivers. We also introduce EyeCar, a new driver visual attention dataset during accident-prone situations. We conduct comprehensive experiments and show that MEDIRL outperforms previous state-of-the-art methods on driving task-related visual attention allocation on the following large-scale driving attention benchmark datasets: DR(eye)VE, BDD-A, and DADA-2000. The code and dataset are provided for reproducibility.