Representation Learning of Pedestrian Trajectories Using Actor-Critic Sequence-to-Sequence Autoencoder