Modelling the unidirectional and bidirectional flow of pedestrians based on convolutional neural networks

In the current urbanisation process occurring worldwide, the management of pedestrian flow in large public spaces to ensure public area safety has become a highly scrutinised issue. This paper introduces a convolutional neural network-based model for simulating pedestrian movement, aimed at improving crowd management and public safety. The model predicts pedestrian trajectories by analysing historical data and comprises four key components: trajectory embedding network, encoder, decoder, and trajectory output network. In addition, the model employs highly parallelisable fully connected layers and convolutional layers to efficiently handle temporal dependencies. The results demonstrate the excellent performance of the model in both unidirectional and bidirectional pedestrian flow scenarios. For example, the model not only successfully reproduced pedestrians' self-organising behaviour (lane formation) but also rapidly (< 5 ms) and accurately simulated their fundamental features, such as density, velocity, and flow rate. To quantitatively evaluate the precision of the simulation, the average displacement error (ADE) and final displacement error (FDE) were applied and were calculated to be 0.104 m and 0.188 m, respectively, in unidirectional flow scenarios, and 0.126 m and 0.226 m, respectively, in bidirectional scenarios. Furthermore, the fluctuation of ADE across various scenarios remained within 0.05 m, and trajectories with ADE exceeding 0.3 m accounted for less than 5 % of the total, demonstrating the model's strong generalisability and robustness. The results indicate that the model is reasonable and capable of rapidly providing situational awareness for security personnel and enhancing crowd management.