The Importance of Loss Functions for Increasing the Generalization Abilities of a Deep Learning-Based Next Frame Prediction Model for Traffic Scenes