Dynamic pedestrian trajectory forecasting with LSTM-based Delaunay triangulation

摘要

Pedestrian trajectory prediction is important for understanding human social behavior. Since the complex nature of the crowd dynamics, it remains a challenging work. Recent studies based on LSTM or GAN have made great progress in sequence prediction, but they still suffer from limitations of modeling neighborhood and handling pedestrian interaction. To address these problems, we propose a conflict-avoiding approach to predict pedestrians’ trajectories based on the Delaunay triangulation graph, which can model the crowd hierarchically. Meanwhile, the middle-level semantic feature is adopted to represent pedestrians’ dynamic interactions in Delaunay triangulation graph. Besides, to evaluate the effect of an additional semantic feature for LSTM, we add an information selection mechanism of pedestrian motion which updates the cell state of LSTM with a new social conflict gate. Furthermore, the results on two public datasets, BIWI and UCY, reveal that the proposed conflict-avoiding approach is excellent in terms of stability and validity. Our experimental results demonstrate that our method can predict the same time span using shorter observation period than state-of-the-art algorithms.