应交通科学与工程学院付川云副教授邀请，加拿大英属哥伦比亚大学Rushdi Alsaleh研究员于12月2日（周五）举行“国际学者云课堂”线上学术讲座，题目为“Multi-agent Modeling of Motorcyclist and Pedestrian Conflicts: Understanding and Simulating Behavior in Mixed Traffic using Machine Learning”，欢迎感兴趣的师生参加。
讲座地点：腾讯会议ID 172-885-711 密码：202255
Road safety evaluation of vulnerable road user interactions (e.g., motorcyclists, pedestrians) in mixed traffic environments is important for promoting active modes of transportation. Safety evaluations have often been conducted using microsimulation models by calculating the number of conflicts based on simulated road user trajectories. However, these models present several shortcomings that might limit the accuracy of the safety evaluation. First, these models are based on several parameters that might influence driving behavior, and small changes in these parameters can significantly change the results. Second, most microsimulation models have been developed under pre-specified rules that tend to avoid crashes, which makes it challenging to represent the actual behavior in conflict situations (e.g., evasive actions). Finally, the models that account for the behavior of vulnerable road users are still limited. This research proposes a Markov-Decision-Process to model the interactions between motorcyclists and pedestrians in conflict situations. This research uses Inverse Reinforcement Learning to recover the reward function from actual trajectories of conflict situations. These trajectories were extracted from a busy and congested mixed-traffic location in Shanghai, China, using computer vision techniques. The reward function is then used to predict the agents’ policies using Actor-Critic Deep Reinforcement Learning. Finally, this research considers the multi-agent framework, where both agents’ are modeled simultaneously. Unlike the traditional single-agent modeling, multi-agent modeling can account for the equilibrium in the agents’ intentions. Results show that the multi-agent model outperformed the single-agent model considering the accuracy of the predicted evasive actions (i.e., changes in speed and direction) and road user trajectories. Moreover, the multi-agent model provided reasonably good predictions of the Post-encroachment time (PET) conflict indicator. In addition, as there are several concerns regarding the interaction between autonomous vehicles and vulnerable road users, models such as the proposed in this research can be used to simulate these emerging environments.
Dr. Rushdi Alsaleh joined the department of Civil Engineering as a postdoctoral research fellow in December 2021. Dr. Alsaleh obtained his Ph.D. in Civil and Environmental Engineering (Transportation Engineering) from the University of British Columbia (Vancouver) in 2021. He has participated in several consulting projects in traffic safety in North America and internationally. Dr. Alsaleh’s research has been focused on improving road safety analysis and evaluation techniques, developing novel techniques for analyzing and modeling active road user in shared spaces using recent advances in machine learning (ML), improving the current knowledge of the safety implications of highway designs, and developing and evaluating Intelligent Transportation System (ITS). He has published more than 20 peer-reviewed journal papers (e.g., Transportation Research Part C & F, Transportmetrica A & B, Transportation Research Record, Accident Analysis and Prevention, Expert Systems with Applications).