Learning-based Microscopic Pedestrian Dynamics Models Based on Motion Interaction Field

Abstract

Research on pedestrian and evacuation dynamics is essential for understanding crowd behavior, optimizing architectural design, and enhancing public safety. Traditional rule-based models can reproduce specific pedestrian self-organization phenomena but often fall short in terms of realism and precision when simulating microscopic behaviors, results in simulations cannot reflect the macroscopic motion characteristics of pedestrians that are in line with reality in engineering applications. Recent years, with the emergence of data-driven approaches, machine learning and deep learning techniques have introduced new perspectives to pedestrian dynamics research. Their robust feature extraction and nonlinear fitting capabilities significantly enhance the accuracy and realism of pedestrian behavior modeling.

The study presented in this thesis focuses on pedestrian behavior modeling methods in pedestrian and evacuation dynamics, systematically exploring the potential applications of intelligent algorithms in this field. Firstly, the concept of motion interaction field is proposed to obtain formatted input data required for intelligent algorithms, and the performance differences between two typical machine learning algorithms are explored based on algorithm characteristics; Subsequently, the motion interaction field was expanded for bottleneck and turning scenarios. By introducing the "herd effect" and scene perception layer, specific intelligent pedestrian behavior modeling methods for different scenarios were proposed based on algorithm characteristics and performance differences. The effectiveness of the established model simulation results was verified from multiple levels; Finally, a discretized expression of the motion interaction field was established for cross scenario applicability and pedestrian collision resolution problems. A general matrix to matrix modeling method and network framework were proposed, and a transferable pre-training model was constructed. By combining transfer learning techniques with controllable experimental data from different scenarios, the transfer layer parameters of the network were fine-tuned, verifying the multi scenario adaptability and generalizability of the modeling method. The main research content is summarized as follows:

(1) Establishment of basic motion interaction field for data-driven modeling of pedestrian microscopic velocity and performance evaluation of typical machine learning algorithms:
This study explores the actual factors influencing pedestrian movement characteristics, constructs a Motion Interaction Field to extract input features for the model, and proposes a data extraction method to address inconsistencies in feature dimensions. Representative intelligent algorithms were selected for training, and the models were subjected to dynamic testing. Performance comparisons of different algorithms were conducted through quantitative and qualitative analyses during both training and testing phases. The reasons for performance differences were examined from the perspective of algorithm characteristics. The findings indicate that decision tree algorithms achieve better prediction accuracy under low- and medium-density conditions, while neural network algorithms demonstrate stronger robustness in high-density scenarios. This work not only verifies the superiority of data-driven methods over traditional social force models but also clarifies the applicable scenarios of different algorithms, thereby laying a methodological foundation and providing a basis for algorithm selection in subsequent research.

(2) Introduce pedestrian heard effect to basic motion interaction field and development of high-density bottleneck scenario pedestrian microscopic behavior modeling method:
Using controlled experimental data of pedestrians passing through a bottleneck, this study incorporates the herd effect to develop a sub-model for predicting movement direction, which determines the actual orientation of the Motion Interaction Field. A second sub-model for predicting speed magnitude was trained using features extracted from the direction-aligned Motion Interaction Field. Together, these sub-models simulate pedestrian flow in front of a bottleneck, with effectiveness validated through both macroscopic characteristics and microscopic error analysis. External test data were introduced to evaluate the potential of the proposed modeling approach for broader applications under limited training data conditions. This research achieves accurate simulation of complex pedestrian flow (featuring both directional differentiation and speed variation) in front of a bottleneck, demonstrating the effectiveness of integrating the basic Motion Interaction Field framework with pedestrian herd behavior, representing a key step toward applying the foundational method to typical complex structures.

(3) Propose a scene perception-motion interaction collaborative network based on two-dimensional motion interaction field and implement modeling of pedestrian turning behavior considering real-time scene perception:
By incorporating a scene information encoding module into the Motion Interaction Field, its prototype was extended to a two-dimensional form to output pedestrian velocity vectors. A scene perception layer and a motion interaction layer were defined to extract environmental information and interactions among pedestrians, respectively. A convolutional neural network was employed to process multi-dimensional input features, forming a scene perception–motion interaction collaborative network. The model was trained using controlled experimental data of pedestrians passing through a turning corridor and demonstrated advantages over the social force model in dynamic testing in terms of fundamental diagrams and headway distances. Finally, the model’s generalization capability was tested using controlled experimental data from a straight corridor. This study enables the extended Motion Interaction Field to perceive the environment and make real-time decisions, thereby enhancing the model’s ability to handle complex scenarios.

(4) Discretize the motion interaction field and combine transfer learning techniques to construct pre-training model and establish multi scenario generalization performance verification:
A discretized representation of relative distances and neighbor velocities within the Motion Interaction Field was established, and scene information was integrated to construct an input matrix. The desired pedestrian motion states were discretized into a motion state space as the output matrix. Considering the characteristics and performance differences of typical intelligent algorithms, a matrix-to-matrix prediction network framework was developed. A pre-trained model was trained using straight corridor experimental data, and parameters in the transfer layer were fine-tuned using data from bottleneck, turning, and real-world scenarios. Testing based on pedestrian movement indicators demonstrated that the transferred model achieves collision-free simulation across various scenarios while ensuring consistency with real pedestrian motion. Its predictive performance comprehensively surpasses that of traditional models, ultimately forming a unified solution for pedestrian behavior modeling that is applicable and scalable across multiple scenarios.

In summary, this thesis focuses on machine learning–based pedestrian and evacuation dynamics modeling and develops the research following a progressive logic of “fundamental methodology development – typical scenario extension – multi-scenario generalization.” First, the fundamental concept of the motion interaction field is proposed, and the modeling performance of representative machine learning algorithms is evaluated, providing a unified input representation and methodological foundation for data-driven pedestrian behavior modeling. Subsequently, for two typical scenarios, bottleneck and turning, the motion interaction field is further extended by incorporating herding behavior mechanisms and scene perception information, and corresponding intelligent models are developed to achieve accurate simulations of pedestrian movement behaviors in these scenarios. Finally, by discretizing the representation of the motion interaction field and integrating transfer learning techniques, a unified matrix-to-matrix network framework and a transferable pre-trained model are constructed, enabling the model to generalize across multiple scenarios. The results demonstrate that the proposed pedestrian behavior modeling approach not only overcomes limitations of previous studies in handling complex environments and interaction information, but also significantly improves the realism and general applicability of microscopic pedestrian behavior simulations. The proposed framework provides a new intelligent solution for pedestrian safety–related applications, such as building fire safety design and evacuation route optimization.

Date of Award	17 Apr 2026
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Lizhong Yang (External Supervisor), Kwok Kit Richard YUEN (Supervisor) & Wai Ming LEE (Co-supervisor)