Predicting Pedestrian Crossing Behavior at Unsignalized Mid-Block Crosswalks Using Maximum Entropy Deep Inverse Reinforcement Learning

In the future driverless scenario, pedestrian-vehicle conflict is an unavoidable traffic problem at unsignalized mid-block crosswalks. Autonomous vehicles are obviously impossible to make eye contact with pedestrians the way a human driver can. Therefore, there is an urgent need for autonomous vehicles to achieve accurate predictions of pedestrian crossing behavior. In order to better solve the problem of high uncertainty of pedestrian crossing behavior, in this paper, a modeling framework combines Maximum Entropy Deep Inverse Reinforcement Learning (Deep MEIRL) and reinforcement learning is employed to predict pedestrian crossing behaviors. The dataset of drone-based video footage is collected in Xi'an (China) to train and validate the model. The trajectory dataset, extracted by computer vision algorithm, is implemented to derive state features of pedestrian behavior, including pedestrian to target area distance, vehicle type, vehicle speed, lateral/longitudinal distances, pedestrian lateral/longitudinal velocities. The results reveal that Deep MEIRL performs better than a baseline model MEIRL at micro-scales such as predicting pedestrian trajectories and evasive actions. Specifically, in predicting pedestrian trajectories, the Deep MEIRL outperforms the MEIRL by 33.09% and 15.16% on the basis of the MAE and HD, respectively. Meanwhile, the Deep MEIRL is 28.7% and 17.6% respectively more accurate than the MEIRL in predicting pedestrian evasive actions on lateral and longitudinal directions. Furthermore, we also found that there is heterogeneity in pedestrian crossing behavior when interacting with different vehicles. This research can contribute to a critical step toward addressing the safe and efficient movement at unsignalized mid-block crosswalks for autonomous vehicles.