The rapid acceleration of urbanization has led to higher concentrations of crowds in confined spaces including high-rise, underground, and large public structures. Especially in emergencies such as fires, the efficiency of evacuation procedures becomes critical for safeguarding human life. Initially, traditional pedestrian and evacuation (PED) models such as Social Force Model (SFM) and Cellular Automaton (CA) had brought insights inform the design of more efficient evacuation facilities and contribute to the development of effective emergency response protocols. However, the effectiveness of simulation conducted by traditional models is disputed due to the assumed rule defined to drive the pedestrian motion.

The model establishing and process. (a) The interaction data encoding module considering ‘herd effect’ and social distance (2.1m); (b) The simulated global flow versus exit width diagram from simulation of bottleneck scenario; (c) The deep neural network with two functional layer which combining the scene perception and motion dynamics of simulated pedestrian; (d) The evaluated quantitative prediction error – trajectory error and duration error from simulation of turning scenario (0.148 m and 0.146 s, respectively).

The research team of Prof. Lizhong Yang, from SKLFS, established novel PED modeling methods by incorporating data-driven approaches from the field of artificial intelligence. By encoding the ‘herd effect’ and scene perception module into the machine learning algorithms, learning-based PED models were trained using data from controlled pedestrian experiments. With the bless of real-world data participating in modeling phase, the simulated results of proposed models have been qualitatively and quantitatively evaluated to be those of superior to traditional models, and such successful trial will bring new perspectives to the study of PED models. The associated works were published in the IEEE Transactions on Intelligent Transportation Systems (24 (2023):7035-7047 and doi.org/10.1109/TITS.2025.3526183) and Expert Systems with Applications (262(2025): 125706).