Pedestrian Trajectory Prediction for Robot Navigation in Crowds

The thesis reports on a data-driven human trajectory prediction model to support robot navigation in environments shared with humans and cluttered with static obstacles. To be able to navigate in such environments, a robot has to predict the movement of the humans around him, as well as navigate in a way that humans can predict the movement of the robot. Therefore a data-driven approach is proposed here. Human trajectories are influenced by their goal, other humans in their vicinity and by static obstacles in their local surroundings. The thesis introduces prediction models based on Long Short-Term Memory models, which use pedestrian movements and an occupancy map of static obstacles as inputs. Different models differentiate themselves in their general architecture as well as in their training methods. The goal of the models is to predict multiple trajectory steps of real world data. Furthermore, the prediction models should generalize well, such that they can be used to predict trajectories of data unknown to them. The various models are used to build a newly developed ensemble model, which is able to predict the majority of all possible trajectories accurately, while being computationally efficient. The selection method of the models used in the ensemble aims on minimizing the risk of collision during navigation as well as maximizing efficiency of the navigation. An extensive evaluation on recorded real-world data is presented, which shows that all models have comparable performance to state-of-the-art models and that the ensemble model increases the number of predicted trajectories drastically. The evaluation is conducted on the basis of several introduced metrics, all adjusted to trajectory prediction