A team at The Korea Advanced Institute of Science and Technology (KAIST) has developed a new AI system aimed at preventing crowd crush incidents by anticipating dangerous densities not just through headcounts, but by analyzing movement flows as well, tells Tech Xplore. Traditional methods tend to treat crowd risk as a function of how many people are concentrated in a space, but that often misses key cues about where and how people are moving. The KAIST team argues that those dynamics matter just as much in detecting imminent danger.
Their approach models crowds using a time-varying graph: “node information” captures how many people are in a particular area, while “edge information” describes how people move between zones. By combining both, the AI can detect early signs of congestion that simple headcounts would miss; for instance, anticipating that a nearby alley will overflow based on inbound flow.
To train this system, the researchers applied a form of 3D contrastive learning, which captures spatial and temporal relationships in the data. They also created six real-world datasets from cities such as Seoul, Busan, Daegu, and New York, and public health sources (e.g., COVID transmissions) to ground their model in diverse movement patterns. On benchmarks, their method outperformed prior approaches by up to 76.1% in prediction accuracy. That margin suggests real potential: at large events or in urban settings, this AI could flag zones likely to become dangerous before they reach critical levels.
The researchers see multiple applications; beyond crowd safety at events, the system could help manage traffic, improve evacuation planning, or even monitor disease spread in dense areas. Their hope is that blending movement analysis with density modeling will transform crowd monitoring from reactive response to proactive prevention.