Multi-Agent-Modellierung der Dynamik von dichten Fußgängermengen: Vorhersagen & Verstehen

• Antragsteller: Mohcine Chraibi, Ph.D.; Professor Dr. Antoine Tordeux
• Fachliche Zuordnung: Verkehrs- und Transportsysteme, Intelligenter und automatisierter Verkehr
• Förderung: Förderung von 2020 bis 2024
• Projektkennung: Deutsche Forschungsgemeinschaft (DFG) - Projektnummer 446168800

• Erstellungsjahr: 2024

Zusammenfassung der Projektergebnisse

In the following paragraphs, we provide a comprehensive overview of the key outcomes of the project. The focus spans various aspects of pedestrian trajectory prediction, deep learning methods, hybrid model developments, trajectory analysis frameworks, crowd dynamics classification systems, and insights derived from a notable field data collection initiative during the Festival of Lights in Lyon. Additionally, we address challenges encountered during the project, including those posed by Covid-related restrictions and unexpected events, alongside the public interest garnered by our research efforts. Deep Learning and Knowledge-based Models Comparison: The project emphasised current trends in pedestrian trajectory prediction using deep learning algorithms, in particular pioneering work such as Social LSTM and Social GANs. It conducted a comparative empirical analysis between the deep learning algorithms and traditional knowledge-based models, such as the Social Force Model and Optimal Reciprocal Collision Avoidance, with the aim of developing hybrid approaches. In addition to classical displacement metrics, new metrics were developed to evaluate and compare the predictions, in particular collision-based metrics. Extended Social LSTM Algorithm: An enhanced version of the Social LSTM algorithm incorporates a timeto-collision loss function. Another hybrid approach based on the LSTM is a two-stage algorithm. It consists of first classifying the scene according to the density, before performing the prediction using networks trained for each density level. Such approaches allowed us to significantly reduce the displacement errors and the presence of collisions in the predictions. Trajectory-Based Analysis Framework: This part of the project developed a trajectory analysis framework using multiple indexes (macroscopic, microscopic, distribution, time series) and methods such as the Kolmogorov- Smirnov test and dynamic time warping. This framework allowed quantifying similarities and evaluating performance in various pedestrian dynamics models. Novel Classification System for Crowd Dynamics: The project introduced two dimensionless numbers related to pedestrian psychology, which helped to understand and classify different crowd dynamics. This novel system, validated with extensive empirical data, improves model validation by identifying different behaviours and crowd flow regimes. Collection of Field Pedestrian Trajectories during the Festival of Lights in Lyon: As part of the project, we were able to collect and make public more than 7,000 pedestrian trajectories with an average duration of 12 seconds and densities of up to 3 pedestrians per square meter. This collection campaign was carried out during the Festival of Lights in Lyon in 2022, using video recordings from cameras previously installed by the project team, as well as high-resolution surveillance cameras from the city of Lyon in cooperation with the technical organisation committee of the Festival of Lights and the project management department of the City of Lyon. The imposition of Covid-related restrictions across Europe posed significant hurdles to direct interactions and staff exchanges. Furthermore, collecting data in dense areas proved to be a challenging task due to a variety of external elements, such as rainfall and the 2022 FIFA World Cup games, which disrupted the activities of the Light Festival.

Projektbezogene Publikationen (Auswahl)

• A generalized trajectories-based evaluation approach for pedestrian evacuation models. Safety Science, 147, 105574.
  Xiao, Yao; Xu, Jun; Chraibi, Mohcine; Zhang, Jun & Gou, Chao
  
• Heterogeneity-induced lane and band formation in self-driven particle systems. Scientific Reports, 12(1).
  Khelfa, Basma; Korbmacher, Raphael; Schadschneider, Andreas & Tordeux, Antoine
  
• Review of Pedestrian Trajectory Prediction Methods: Comparing Deep Learning and Knowledge-Based Approaches. IEEE Transactions on Intelligent Transportation Systems, 23(12), 24126-24144.
  Korbmacher, Raphael & Tordeux, Antoine
  
• Come together: A unified description of the escalator capacity. PLOS ONE, 18(3), e0282599.
  Gnendiger, Christoph; Chraibi, Mohcine & Tordeux, Antoine
  
• Single-File Pedestrian Dynamics: A Review of Agent-Following Models. Modeling and Simulation in Science, Engineering and Technology, 143-178. Springer International Publishing.
  Cordes, Jakob; Chraibi, Mohcine; Tordeux, Antoine & Schadschneider, Andreas
  
• Time-Continuous Microscopic Pedestrian Models: An Overview. Modeling and Simulation in Science, Engineering and Technology, 55-80. Springer International Publishing.
  Korbmacher, Raphael; Nicolas, Alexandre; Tordeux, Antoine & Totzeck, Claudia
  
• TTC-SLSTM: Human Trajectory Prediction Using Time-to-Collision Interaction Energy. 2023 15th International Conference on Knowledge and Systems Engineering (KSE), 1-6. IEEE.
  Dang, Huu-Tu; Korbmacher, Raphael; Tordeux, Antoine; Gaudou, Benoit & Verstaevel, Nicolas
  
• Dimensionless numbers reveal distinct regimes in the structure and dynamics of pedestrian crowds. PNAS Nexus, 3(4).
  Cordes, Jakob; Schadschneider, Andreas & Nicolas, Alexandre
  
• Predicting pedestrian trajectories at different densities: A multi-criteria empirical analysis. Physica A: Statistical Mechanics and its Applications, 634, 129440.
  Korbmacher, Raphael; Dang, Huu-Tu & Tordeux, Antoine
  
• Toward Better Pedestrian Trajectory Predictions: The Role of Density and Time-to-Collision in Hybrid Deep-Learning Algorithms. Sensors, 24(7), 2356.
  Korbmacher, Raphael & Tordeux, Antoine