Evaluation of Output Representations in Neural Network-based Trajectory Predictions Systems

This work deals with the challenging problem of pedestrian trajectory prediction, when observations from these pedestrians can be gathered through a urban video monitoring system. Since most of state-of-the-art systems in this field are now based on deep recurrent neural networks, here we study one specific characteristic of these systems, namely the way they encode their output. We compare three different representations of the output, and show that those representations working on residuals (in particular, displacements with respect of last pedestrian position or linear regression models of residual errors) produce much more accurate predictions than those ones handling absolute coordinates.