Imitation Learning for Autonomous Vehicle Driving: How Does the Representation Matter?