Pedestrian motion in simulation applications using deep learning

Abstract

The goal of this paper is to provide a framework for simu-lating pedestrian motion in simulation applications by using real-world examples of human motion. This process has two implications. The first one refers to the reduction of the development time, since a deep learning model can replace the classical pedestrian behavior development process for the targeted applications. The second relates to improving the quality of pedestrian movements, as manual development of behavior using classical methods can result in movements that appear too robotic or predictable. We propose a new deep learning model based on an encoder-decoder strategy and Graph Attention Networks, able to take into account both the semantics of the scene and the correlations between the simulated pedestrian movements. The evaluation shows that the methods are suitable for real-Time simulations, even for applications with performance constraints such as video games.