Machine Learning and AI Help NJIT Researchers Understand Human Crowd Movement
A dose of artificial intelligence is helping New Jersey Institute of Technology researchers make sense of how crowds and the individuals within them move around, leading to insights with applications in fields such as emergency management, pedestrian traffic planning, robotics, special effects and even videogames.
Tomer Weiss, assistant professor of Informatics in NJIT’s Ying Wu College of Computing, leads the research with graduate students to see if AI can complement their understanding of the movement patterns in useful ways.
“It is very difficult to come up with rules that realistically capture how crowds behave — there is no Newton law such as F=ma that can tell you exactly what will happen. But, it is very important to understand how crowds move, for example to prevent stampedes in mass events,” Weiss stated.
Traditional machine learning, and what is now largely considered AI, hypothetically allows computation of rules from data. “Rules which the AI model learns cannot be written down as F=ma, but can be represented with a black box that learns crowd motion patterns. Such patterns allow us to gain more insight for modeling crowd behaviors,” Weiss explained.
His team’s software creates thousands of parallel AI agents that represent humans walking through crowds. Over time, the agents learn to interact with each other and move accordingly. They may form groups or change direction. The software uses the popular AI concept of deep reinforcement learning, which in this application commands the agents to walk toward goal locations or find certain locations in the most efficient paths, while avoiding collisions with each other.
Weiss said that in some cases, the AI model performs better than human-designed rules. In a recent paper with graduate students Bilas Talkukdar and Yunhao Zhang, presented at an ACM Siggraph conference, “We demonstrated the robustness of our method in multiple crowd navigation scenarios,” the authors wrote.
Looking forward, the team cited several directions for their research. One direction would be simulating pedestrian scenarios that showcase diverse agent behaviors, such as walking slowly or rushing. Dense crowd flows are another direction, such as how people move when exiting a stadium or during an emergency. Finally, they said that using new methods such as simulated vision — what the virtual agents would see in front of them — may help discover novel crowd behaviors.