Multi-Agent modelling of dense crowd dynamics: Predict & Understand – MADRAS

Trustworthy models for the dynamics of dense crowds are crucial for the prediction of pedestrian flows and the management of large crowds, but also from a fundamental perspective, to understand the roots that they share with active matter but also the pedestrian specifics. However, current models suffer from some severe deficiencies, especially at high density. In this context, MADRAS aims to develop innovative agent-based models to predict and understand dense crowd dynamics (from 2 to 8 ped/m2) and to apply these models in a large-scale case study. Two complementary modelling approaches will be pursued:

(i) neural networks (NN) that will be trained on available data to predict pedestrian motion as a function of their local environment and trajectory. This data-based approach is bolstered by recent successes, which proved the potential of recurrent NN at low to intermediate density, but suitable descriptors for the agent's neighbourhood and the local geometry must be found to address dense crowds in complex geometries.

(ii) a physics-based model coupling a decisional layer, where a desired velocity is selected according to an empirically validated collision-anticipation strategy, and a mechanical layer, which takes care of collisions and contacts. To push this approach to higher densities, integrating more realistic pedestrian shapes and better splitting the decision-making process from mechanical forces is necessary.

These approaches will be confronted with novel validation methods, using data from controlled experiments. The models will then be exploited at larger scale to simulate the flows on crowded streets at a real mass gathering, the Festival of Lights in Lyon. To this end, empirical data will be collected by filming the streets from above and by immersing in the crowd participants wearing pressure-sensing jackets, to measure contacts.

Emulating this real scenario will call for adequate data assimilation methods and efficient multi-agent simulations based on the two models. The latter will be combined in a single online platform, allowing one to visualise the predicted flows and compare them with the ground truth.

Finally, the impact of different model ingredients and features (e.g., shape heterogeneities) on the large-scale flow predictions will be investigated by means of numerical simulations of the two models, using the Festival of Lights situation as reference scenario.

To achieve this ambitious interdisciplinary project, four teams with different backgrounds (Computer Science, Statistical Physics, Applied Mathematics) will combine their strengths and research tools (simulation platform, continuous models, datasets, experimental analysis tools). Cooperation will be fostered by the common goal to reproduce a large-scale scenario and by extensive (3-month) research stays at another institute for the involved PhD students.