Deep Learning for Predicting Pedestrian Trajectories in Crowds

In the last decade, deep learning algorithms have gained a lot of attention in many fields. Even scientists who study pedestrian behavior and usually create physics-based models to simulate realistic behavior are increasingly using these algorithms. So far, however, this has only been tested in low-density environments. In this contribution, we want to investigate the question of whether the methods are also suitable for crowds. Therefore, we prepared high-density datasets and tried to predict the behaviors of the pedestrian with physics-based models as well as deep learning algorithms. We compare the prediction accuracy in terms of two error metrics: a distance-based metric and a metric that counts the number of collisions between pedestrians. The results show that the algorithms beat the models regarding the distance metric, but perform worst in terms of the collision metric. To be able to predict more realistic crowd behavior, the deep learning algorithm needs to improve collision avoidance. In order to make more accurate predictions, the deep learning algorithms need to be improve collision avoidance.