Core Concepts
The proposed visual-information-driven (VID) model effectively captures visual information, including scenario geometry and pedestrian locomotion, to enhance the adaptability of data-driven crowd simulation models across diverse geometric scenarios.
Abstract
The paper proposes a novel visual-information-driven (VID) model for crowd simulation. The key highlights are:
The VID model emphasizes the importance of incorporating visual information, such as scenario geometry and pedestrian locomotion, to improve the adaptability of data-driven crowd simulation models.
The VID model consists of three modules: a data processing (DP) module, a velocity prediction (VP) module, and a rolling forecast (RF) module. The VP module is a temporal convolutional network (TCN)-based deep learning model named social-visual TCN (SVTCN).
The DP module extracts visual information using a radar-geometry-locomotion (RGL) method, which captures the relative positions of walls and pedestrians around the subject pedestrian.
The SVTCN in the VP module takes the extracted social-visual features and motion data as input and predicts the velocity of the subject pedestrian at the next time step.
Experiments are conducted on three public pedestrian motion datasets with distinct geometries: corridor, corner, and T-junction. Both qualitative and quantitative metrics are used to evaluate the VID model, and the results demonstrate its improved adaptability across all three geometric scenarios compared to previous data-driven models.
Stats
The flow J of the measurement area is defined as J = ρvb, where v, ρ and b denote the Voronoi velocity, density and width of the measurement area, respectively.
The egress time error (ETE) is defined as the absolute difference between the egress times in the simulation and controlled experiment.
The percentage egress time error (PETE) is defined as the ratio of the ETE to the egress time in the controlled experiment.
The travel time error (TTE) indicates the difference between the simulated and actual travel time.
The percentage travel time error (PTTE) represents the TTE as a percentage of the actual travel time.
The trajectory displacement error (TDE) quantifies the mean displacement error of a pedestrian throughout their travel between controlled experiments and simulations.
The final displacement error (FDE) measures the displacement error of a pedestrian at the end of their travel, specifically when they exit the scenario.