Modeling Pedestrian Crossing Behavior At Unsignalized Mid-Block Crosswalks Using Maximum Entropy Deep Inverse Reinforcement Learning

Unsignalized mid-block crosswalks are inevitably shared by pedestrians and autonomous vehicles in the future.Therefore,the possible future behavior of pedestrians must be considered when planning the route of autonomous vehicle.In addition,in order to better understand the existing pedestrian vehicle interaction in this road section,further research and modeling of pedestrian crossing behavior are needed.Compared with other traffic participants,pedestrian crossing behavior has complex and uncertain characteristics,making its modeling challenging.Therefore,this paper proposes an improved modeling framework that introduces the Maximum Entropy Deep Inverse Reinforcement Learning(Deep MEIRL)and the Proximal Policy Optimization(PPO).In this framework,Deep MEIRL learns the reward function of pedestrian crossing behavior from pedestrian trajectory data,PPO generates the optimal pedestrian crossing strategy based on the reward function,and finally predicts pedestrian crossing behavior through this strategy.To adapt to the modeling task of this paper,a customized neural network structure is introduced in the Deep MEIRL.This study used video data captured by drones near Yanta Road and Hepingmen Subway Station in Xi’an,Shaanxi.By employing advanced machine vision algorithms,namely YOLOv5 and Deep SORT,trajectory data of pedestrians and vehicles were extracted from the videos.The specific research work is as follows:Firstly,the range of modeling-related features and training termination conditions was established.Descriptive statistical analysis was performed to examine the state and action characteristics of pedestrians,while employing one-way ANOVA to test the significant influence of vehicle type on pedestrian crossing behavior.The results indicate that there are significant differences in pedestrian crossing behavior when encountering various types of vehicles.Secondly,the objective was to quantify the disparities in pedestrian crossing behavior while engaging with diverse vehicle types,as well as to visualize the reward function generated by the Deep MEIRL neural network.The outcomes reveal that when pedestrians interact with buses,the lateral and longitudinal distances of pedestrian orientation are-1 meter and 3 meters,respectively.In contrast,when pedestrians interact with cars,the corresponding distances are 1meter and 5 meters,respectively.These findings suggest that pedestrians exhibit greater trist on buses than cars when it comes to street crossing.Finally,the proposed modeling framework is used to predict pedestrian crossing behavior and evasive behavior,and its predictive performance is compared with traditional modeling frameworks based on maximum entropy inverse reinforcement learning(MEIRL)and PPO.The results show that the modeling framework proposed in this paper reduces the prediction error by an average of 41.8% when predicting pedestrian crossing behavior,and by an average of 23.2% when predicting pedestrian avoidance behavior.The possible reason is that compared to MEIRL’s linear reward function,Deep MEIRL can fit the nonlinear reward function to learn more microscopic changes in pedestrian behavior,thereby better adapting to complex pedestrian crossing behavior modeling.This paper proposes an improved modeling framework,and considers the difference of pedestrian crossing behavior when facing different types of vehicles.The model can more accurately predict pedestrian crossing behavior and avoidance behavior,which can provide some reference for the trajectory prediction research of autonomous vehicle.