AI Breakthrough Could Prevent Crowd Crush Disasters

Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have developed breakthrough AI technology that could prevent crowd crush disasters by predicting dangerous density patterns before they occur. The innovation combines real-time population data with movement flow analysis to identify potential safety risks at large-scale events.

Context and Background

The research team, led by Professor Jae-Gil Lee from KAIST’s School of Computing, has created predictive AI that analyses both population density and movement patterns simultaneously. Unlike previous approaches that focused on either current crowd size or movement direction separately, this technology combines both factors using a sophisticated time-varying graph system.

The breakthrough achieved up to 76.1% improvement in prediction accuracy over existing state-of-the-art methods. The team validated their approach using real-world datasets from Seoul, Busan, and Daegu subway systems, New York City transit data, and COVID-19 case distribution patterns across South Korea and New York.

The technology employs what researchers term “bi-modal learning” with 3D contrastive learning techniques. This allows the AI to understand not just current crowd distribution, but how populations develop dangerous patterns over time by analysing spatial relationships and temporal changes simultaneously.

Looking Forward

The innovation addresses critical gaps in crowd management technology that became evident after incidents like the Itaewon tragedy. By predicting when specific areas will become dangerously congested—rather than simply counting current occupancy—the system enables proactive safety interventions.

Professor Lee emphasised the technology’s broader applications beyond event management, including urban traffic optimisation and infectious disease outbreak response. The research team has publicly released six real-world datasets to accelerate further development in crowd prediction technology across the global research community.

Source Attribution:

Share this article