Abstract:

In order to explore the safety of interactions between pedestrian groups moving in different directions, a risk assessment method for pedestrian safety was proposed based on a Seq2Seq network integrated with an AM. By analyzing the interactions among pedestrians, local density was introduced as an additional input feature to better characterize dynamic interpersonal behaviors. The observed sequential features of pedestrians within a continuous time window were fed into an LSTM-based encoder-decoder architecture, and AM was employed to capture multidimensional critical information during motion, enabling the reconstruction of complex crowd movement patterns across various scenarios. Furthermore, pedestrian safety risks were quantified by introducing a crowd pressure metric, and the pressure value ranges corresponding to three typical movement states were calculated to enable risk stratification. The results show that the average displacement error (ADE) and final displacement error (FDE) in trajectory prediction are less than 0.3 m under both unidirectional and bidirectional flow conditions at different densities, indicating the model's high accuracy in trajectory prediction. Based on the risk variation observed in bidirectional flow scenarios, it is found that the probability of accidents increases significantly when pedestrians' velocities change abruptly under high-density conditions. Therefore, timely warnings of crowd states and proactive intervention measures are required to mitigate potential safety risks.

Key words: sequence to sequence (Seq2Seq), crowd safety, risk assessment, trajectory prediction, attention mechanism(AM), crowd pressure