Modeling Crossing Behaviors of E-Bikes at Intersection With Deep Maximum Entropy Inverse Reinforcement Learning Using Drone-Based Video Data

Abstract

The crossing behaviors modeling of non-networked road users can improve Connected and Autonomous Vehicles (CAVs)’s awareness of the imminent hazards in shared space while planning routes. In this study, an agent-based microsimulation model is utilized to simulate the crossing behavior of e-bikes which is difficult to model due to their greater maneuverability. To deeply understand the crossing mechanism of e-bikes, a customized neural network-based nonlinear reward function is developed from five dimensions, including travel efficiency, travel direction, travel destination, risk avoidance, and travel location. A Deep Maximum Entropy Inverse Reinforcement Learning(Deep MEIRL), which can recover the nonlinear reward function, is introduced to predict trajectories of e-bikes from the real drone-based video dataset, collected at the intersection of Jixiangcun, Xi’an (China). The results reveal that the Deep MEIRL can simulate the e-bike crossing trajectory more precisely, particularly in the microscopic behaviors of riders. Comparing Deep MEIRL with the baseline model MEIRL, it can be found that the nonlinear reward function designed in this paper is more advantageous in terms of continuous space modeling, with an improvement of 19.77%. Notably, Deep MEIRL outperforms MEIRL in modeling distance to potentially risk target states for left-and right-turn crossing behavior, improving Mean Absolute Error (MAE) by 71% and 30%, respectively. It means that the Deep MEIRL with a designed neural network is of great value to provide a logical result in modeling e-bike interaction with other road users. Therefore, the research is conducive to CAVs to understand e-bike behaviors in complex traffic scenarios, thereby assisting CAVs in making decisions efficiently.