Abstract :
This paper addresses the problem of virtual pedestrian autonomous navigation for crowd simulation. It describes
a method for solving interactions between pedestrians and avoiding inter-collisions. Our approach is agent-based
and predictive: each agent perceives surrounding agents and extrapolates their trajectory in order to react to potential
collisions. We aim at obtaining realistic results, thus the proposed model is calibrated from experimental
motion capture data. Our method is shown to be valid and solves major drawbacks compared to previous approaches
such as oscillations due to a lack of anticipation. We first describe the mathematical representation used
in our model, we then detail its implementation, and finally, its calibration and validation from real data.
