Neural network model for predicting variation in walking dynamics of pedestrians in social groups

Pedestrian spaces are increasingly becoming popular locations for shopping, recreation, festivities, and other social activities. Therefore, an improved understanding of the factors that make walking environments enjoyable and safe is essential. Most existing studies focus on modelling walking behaviours of individual pedestrians. However, most people participate in these activities as parts of social groups. Although the movement and choice behaviours of pedestrians in social groups differ from those of individuals, a model featuring group movements has not been developed. This study uses neural networks to analyse the effects of variables influencing pedestrian movements of social groups and predict the variation in walking dynamics. A top-view video was used to extract the trajectories of pedestrian groups. After identifying the social groups in a crowd, the movement characteristics, pedestrian–environment interaction, inter-pedestrian interaction, intra-group relationship, and inter-group relationship of all group members were calculated and considered in the model. After a variable selection process using neural networks, a neural network model was developed featuring variables that are strongly related to the lateral or longitudinal changes in the individual’s walking speed. The current movement condition, presence of obstacles nearby, impending collisions, current position and velocity of other group members, and following behaviour were found to impact a pedestrian’s walking dynamics. The proposed model can predict the pedestrian density and distribution according to a space function, contributing to better crowd management and efficient design and renovation of pedestrian spaces. Furthermore, the variable selection method can optimise and simplify other pedestrian behaviour prediction models.