KAIST Develops AI Technology to Improve Accuracy of Crowd Density Prediction

A new technology has been developed that significantly improves the accuracy of crowd density prediction using artificial intelligence (AI). This technology can be used to prevent mass crowd accidents like the Itaewon disaster, alleviate urban traffic congestion, and respond to the spread of infectious diseases.

KAIST announced on the 17th that a research team led by Professor Jaegil Lee from the Department of Computer Science has developed a novel AI technology capable of accurately predicting crowd density situations.

Research team led by Professor Jaegil Lee, Department of Computer Science. Provided by KAIST

The way crowds gather in a particular space at a specific time cannot be explained simply by the increase or decrease in the number of people. Even with the same number of people, the risk level varies depending on where the crowd is coming from and in which direction they are dispersing.

However, previous studies mostly focused on a single type of information, such as "How many people are currently gathered?" or "Which routes are people flocking to?"

In contrast, the research team focused on the idea that combining both types of information can help detect actual risk signals that may arise in the field.

They also introduced the concept of a "time-varying graph" to represent these movements. By simultaneously analyzing how many people are in a specific area (node information) and how the population flows between areas (edge information), the technology increases the accuracy of predicting crowd density.

For example, a sudden surge in crowd density in alley A is not predicted solely by the current number of people present. By simultaneously identifying the flow of people moving from nearby area B toward A (edge information), the system can detect early warning signals that "area A may soon become dangerous."

To implement this principle, the research team developed a "bi-modal learning" approach. This technology enables AI to simultaneously consider population size (node information) and population flow (edge information), while learning both spatial relationships (connections between locations) and temporal changes (when and how movement occurs).

In addition, the team introduced a 3D contrastive learning technique, allowing the AI to learn "three-dimensional relationships" that incorporate not only two-dimensional spatial (geographic) information but also temporal information.

Through this, the AI can analyze not just whether the population is currently high or low, but also the patterns of crowding over time, enabling it to predict the location and timing of congestion with greater accuracy than before.

The research team also built and released six research datasets by collecting and processing real-world data from the Seoul, Busan, and Daegu subway systems, New York transportation data, and COVID-19 case numbers in Korea and New York. Using these datasets, the accuracy of crowd density prediction improved by up to 76.1% compared to previous methods.

Meanwhile, the study was conducted with Youngeun Nam, a doctoral student at the KAIST Department of Computer Science, as the first author, and Jihae Na, a doctoral student, as a co-author. The research results were presented last month at the international conference on data mining, the Knowledge Discovery and Data Mining (KDD) 2025.