Humans navigate complex environments in an organized yet flexible manner,\nadapting to the context and implicit social rules. Understanding these\nnaturally learned patterns of behavior is essential for applications such as\nautonomous vehicles. However, algorithmically defining these implicit rules of\nhuman behavior remains difficult. This work proposes a novel self-supervised\nmethod for training a probabilistic network model to estimate the regions\nhumans are most likely to drive in as well as a multimodal representation of\nthe inferred direction of travel at each point. The model is trained on\nindividual human trajectories conditioned on a representation of the driving\nenvironment. The model is shown to successfully generalize to new road scenes,\ndemonstrating potential for real-world application as a prior for socially\nacceptable driving behavior in challenging or ambiguous scenarios which are\npoorly handled by explicit traffic rules.\n