Large-scale Crowd Simulation Based on Real-world Data

You are here

Home » Node

Large-scale Crowd Simulation Based on Real-world Data

Managing crowds in a large indoor facility in face of uncertain crowd behaviors is challenging yet of critical importance in ensuring operation safety and efficiency. A large-scale agent-based crowd simulation could be used to simulate realistic crowd behaviors under various management policies that are otherwise difficult to evaluate.

The simulation platform is built using crowd movement models derived from analyzing mobility traces collected via Wifi localization technology. By mining massive amount of crowd trajectories, we are able to derive highly realistic crowd movement models to create a realistic crowd simulation model according to types of events. The simulator provides venue operators with a decision support tool to design and implement efficient egress strategies as part of their day-to-day operation to enhance visitors experience.

In addition, we utilize video analysis from multiple CCTV cameras covering different areas of the venue to extract finer crowd features such as coherent groups (segments), and flow rates and crowd densities. Such crowd features enable us to model realistic crowd behaviors with further details to supplement the general movement models based on the Wifi data.

Our technology is developed and tested at a well-known convention center in Singapore. The crowd simulation platform can be customized to work with other facilities, if required data is available.
 

The crowd simulation platform has the following major innovations:

  • Crowd behaviors are derived directly from raw in-door movement traces, which are sparse, incorrect and noisy when compared to outdoor GPS sensing. To address this issue, we have developed and tested a novel trajectory mining algorithm that can combine a large number of “similar” trajectories to infer the likeliest routes for visitors visiting different parts of the building.
  • Crowd simulation is 3D. To make our simulator easily adoptable, we have designed a unique set of tools that can create a 3D building model based on annotated 2D floorplans, with only minor manual modifications. This design significantly reduces the model building time.
  • By utilizing machine-learning techniques together with the multivariate data (i.e. Wifi trace, CCTV camera videos), we propose a novel method to correlate the source and destination choices and the route choices of crowd to certain layout and motion features that are transferrable to different scenarios. With multiple datasets, the trained weights of such features can thus be applied to simulate and predict crowd movements in new scenarios.
The complexity of configuring and setting up the simulation is hidden from the users with a user-friendly web portal.

 


The primary application area of our technology is in crowd management within a large in-door facility. This may include but not limited to convention and meeting centers, sports stadium, or shopping mall.

The strength of our crowd simulation platform is the ability in satisfying end-to-end modeling and usage needs: The crowd behaviors, which are of critical importance in correctly evaluating the management policy, are directly sensed and derived from the collected movement traces. Our platform can work even when the collected information is of inferior quality. The simulation, once equipped with learned crowd movement models, can also be easily operated using a user-friendly web portal. Through this portal, users can easily adjust parameters to be tested, set up simulation experiments, and collect and observe the generated outcomes.

 

 

SUBSCRIBE TO OUR NEWSLETTER

Keep up to date with what's happening at the Singapore Management University