CEE Seminar - Advancing pedestrian mapping and simulation in urban areas using multimodal sensing and generative models

Works of social infrastructure, such as public parks and markets, play a crucial role in fostering community wellbeing. Understanding pedestrian movement within these spaces is critical for designing effective urban infrastructure. However, traditional approaches to monitoring and simulating pedestrians face significant challenges regarding privacy, scalability, and generalizability. This talk presents a series of works that advance pedestrian spatial analysis using multimodal sensing systems and AI. First, we introduce a new privacy-preserving method to localize pedestrians using ground vibrations generated by footsteps. Deep learning models and probabilistic methods are developed to scale the system for multiple simultaneous pedestrians in large outdoor spaces. Second, we explore the potential of using text-to-video (T2V) and image-to-video (I2V) diffusion models as general purpose crowd simulators. State-of-the-art models are benchmarked for the pedestrian simulation task using a method to extract and evaluate 2D multi-agent trajectories directly from generated videos. We will discuss ongoing experiments to reduce hallucinations in model outputs that cause pedestrians to merge or disappear in simulated scenes. Ultimately, these frameworks offer urban planners and stakeholders exciting new tools to evaluate the quality of public urban spaces and simulate realistic social behaviors.