Interactive simulation of local interactions in dense crowds using elliptical agents

We present a practical approach for interactive crowd simulation based on elliptical agents. Our formulation uses a biomechanically accurate pedestrian representation to simulate different local interactions, including backpedaling, side-stepping, and shoulder-twisting. We present an efficient algorithm for local navigation and collision avoidance among multiple elliptical agents using velocity obstacles. Furthermore, we describe techniques to link the orientation of each elliptical agent to its velocity to automatically generate turning and lateral movements. In practice, our approach can simulate dense crowds of hundreds of pedestrians at interactive rates on a single CPU core. We highlight the performance in complex scenarios and validate our simulation results by comparing with real-world crowd videos and experiments.