|This work discusses the navigation of social pedestrian groups in crowded environments. First, we highlight the necessity for including social-group behavior in crowd simulations. We also compare existing group methods against each other. Next, we propose Social Groups and Navigation (SGN) method for steering social groups of agents in planar homogeneous environments with polygonal obstacles.
Our method provides group-specific details on global and local aspects of pedestrian navigation. SGN is inspired by the social-force model of Moussaïd et al. 2010. In addition, it borrows the vision-based collision-avoidance algorithm of Moussaïd et al. 2011. Both original works have been adjusted to improve the ability of the simulated groups to remain social and coherent, while avoiding obstacles and other groups. SGN is flexible and can be coupled with different navigation meshes and global route-planning algorithms. The method is evaluated through extensive experiments.
Results demonstrate the ability of SGN to produce coherent and social-friendly group configurations throughout the simulation. Based on our evaluation metrics, the quality improvement over the works of Moussaïd et al. is significant. When groups of three or four are considered, SGN produces social friendly configurations at a significantly higher rate than the method of of Moussaïd et al. In all tested scenarios, the difference of this rate ranges from 15% to 31% for groups of three. Regarding groups of four, the difference is even greater, ranging from 13% to 53%. When groups of two are considered, the difference is small, ranging from 1% to 4%, but is still statistically significant. Performance results suggest that SGN is capable of simulating in real-time thousands of agents that are organized in small social groups.
This thesis project was conducted as part of a collaboration between Utrecht University and INCONTROL Simulation Solutions. Our proposed method has been integrated both in Pedestrian Dynamics, a crowd flow simulator developed by INCONTROL, and in the crowd simulation framework developed by Utrecht University. This research has been supported by the COMMIT/ project.