Modeling Spatial-Temporal Social Interactions for Pedestrians Trajectory Prediction on Real and Synthetic Datasets

Abstract

Humans moving in crowded spaces adapt their pace or alter their initial path. A variety of sociocultural factors and personal preferences influence these interactions. With the development of autonomous moving platforms that need to share their physical environment with humans, this task has become increasingly valuable. In this paper, we present a way to conveniently evaluate the ability of a model to predict social interactions between pedestrians. In this manner, we conduct experiments on two prediction models: Social long short-term memory (LSTM) and Vanilla LSTM models. We use real datasets by generating synthetic datasets for which we can define the impact of social interactions on the motion of pedestrians. These hand-tailored datasets exclude individual interactions and focus on the interactions between individuals. By comparing the models’ performances on these datasets, we show that while the Social LSTM model can predict social interactions, the Vanilla LSTM model cannot. For our analysis, we introduce evaluation metrics that focus on the interactions between pedestrians, and these metrics go beyond the commonly used average and final displacement error for trajectory prediction. In particular, we analyze the prediction errors in regions of trajectories highly influenced by social interactions. Furthermore, we analyze the collision behavior of the model’s predictions and classify trajectories concerning their degree of non-linearity. This allows us to compare the models’ performances on motions differently influenced by social interactions.