Characterizing vehicle–pedestrian interaction behavior in near misses: Insights from three different cities

• A multi-agent Markov game model is used to measure the degree of cooperation in pedestrian-vehicle interactions. • Inverse Reinforcement Learning is used to capture road users’ utilities. • The study considers data from Boston (US), Cairo (Egypt), and Singapore. • Cairo represented the most cooperative environment, whereas Singapore was the lowest. • The study provides valuable insights into road users’ microscopic interaction behavior in different traffic environments. Improving the safety of vulnerable road users such as pedestrians requires a good understanding of their interaction behavior and their collision avoidance mechanisms in interactions with other road users. Refining this understanding will become even more important in an automated driving environment, where properly representing road users’ evasive actions is required to develop effective collision avoidance systems, especially in mixed and less organized traffic conditions. This study models vehicle–pedestrian interactions using a multi-agent Markov game modeling framework to measure the degree of cooperation as road users interact with each other (e.g., collectively try to avoid a crash). Data from three cities with different traffic environments were used, including Boston (US), Cairo (Egypt), and Singapore. The model adopts an Inverse Reinforcement Learning framework that captures road users’ utilities from their trajectories while accounting for the equilibrium in their actions. Results demonstrate substantial variations in behavior across different cities. For example, Cairo was shown to be the most cooperative environment, whereas Singapore presented the lowest levels of cooperation. Moreover, the level of cooperation is negatively associated with speed variables, which shows that road users were expected to cooperate more when they reduced their speeds. This paper provides valuable insights into road users’ cooperation levels in different environments. This is useful for accurately modeling road users’ actions and incorporating their behaviors in advanced automated driving systems, which should properly reflect local traffic environment conditions.