A Unified Framework for Hierarchical Pedestrian Behavior Generation in Urban Scenario

Zhengming Zhang, Vincent G. Duffy, Mark R. Lehto, Zhengming Ding, Renran Tian

Abstract: Achieving fully autonomous driving systems hinges on the complex task of modeling pedestrian behavior, which is characterized by unpredictability and diverse actions. While significant progress has been made in predicting pedestrian trajectories and actions, the generation of interactive behaviors for realistic simulations remains underexplored. This study introduces an innovative unified framework designed to generate hierarchical pedestrian motion that not only responds to traffic dynamics but also reflects natural human movements. We utilize existing motion forecasting datasets as the high-level trajectory data and enhance this with egocentric datasets for detailed pose information. By synchronizing annotations from bird’s-eye view and ego-view perspectives, we bridge the gap between macro-level paths and micro-level bodily movements, ensuring that the generated behaviors are coherent and integrated. The ultimate goal is to develop a system that enhances the authenticity of autonomous driving simulations and contributes to the overall safety and reliability of autonomous vehicle technologies.