PERception-based CrOwd simuLATION – PERCOLATION

The project PERCOLATION aims at developing next generation crowd simulators. We focus on pedestrian microscopic crowd simulation. Microscopic simulation is based models of local interactions between agents and tries to simulate the emergence of collective behaviors. Generally, the mathematical formulation of local interaction is simple. PERCOLATION proposes a new approach to microscopic crowd simulation by introducing perception-based models. Compared to previous models, our key idea is to mathematically formulate simulation models based on perceptual variables, and especially visual variables. This means that agents are equipped with synthetic vision and that we simulate agent’s reactions to the pixels that an agent perceives. Our goal is to model a perceptual control law as close as possible to the real human’s perception-action loops that drive local interactions. In spite of the complexity raised by such approaches, we expect an increase of realism in simulations, especially at the collective scales (expectations are based on recent results [Ondrej2010]).
Various questions need addressing. First, we need describing these human perception-action loops that drive the interactions in crowds. Then, we have to transform such descriptions into simulation algorithms. Finally, we want to evaluate the level of realism that is reached with such models. We choose to perform experimental observations to describe interactions as well as to perform evaluation. However, dedicated comparison methods need to be designed. Finally, we suggest exploring Virtual Reality technologies to perform experiments on pedestrian behaviors in crowds. Goal is to enable studying complex situations with controlled conditions as well as to develop new evaluation paradigms.
To achieve these goals, PERCOLATION is organized along 3 axes. Axis #1 will build a kinematics dataset on interactions between humans during locomotion (avoidance, following, meeting, etc.). We want to observe single interactions (1 vs. 1) or complex interactions where numerous subjects. We want to observe typical emergent macroscopic structures in data. Axis #2 is about numerical modeling, crowd simulation and validation. We directly use experimental results to formulate models of interactions and also use the kinematics dataset for evaluation purpose. Finally Axis #3 extends both experimental abilities and explores new validation paradigms by exploring the use of Virtual Reality to immerse people in virtual crowds.
PERCOLATION is a multidisciplinary project. Beyond crowd simulation, we use methodologies from Movement Sciences (Axis #1) and Virtual Reality (Axis #3) to meet our objectives. The PERCOLATION team gathers researchers from Computer Science (INRIA Rennes) and Biomechanics (M2S laboratory) to bring together the required expertise. People involved in the project had frequent and fruitful past collaborations. The project also involves young researchers, PhD students, and proposes internships for masters students to develop this new generation of crowd simulators.
Project’s expectations are many. PERCOLATION goes deeper than ever in modeling the details of interactions occurring between individuals in crowds. We detail the perception-action processes that drive them. We expect extending the number of factors that simulators are able to consider (e.g., lighting condition in buildings thanks to synthetic visual perception). This draws interesting connections to Cognitive Sciences. Also, the PERCOLATION team expertise enables starting a Virtual Reality platform for Immersive Crowds. In case of success, this platform could be a major breakthrough, and is promising for experiments on crowd behaviors, new paradigms for crowd simulation evaluation, or an interesting tool to collaborate with Neuroscience. In terms of applications, we think about architecture, treatment of social phobias, or training of safety staff to crowd management.