In this paper we place ourselves in the broad context of autonomous driving and, more precisely, in the context of road crossing decisions by pedestrians. Machine learning models built by the existing work need as training data datasets constructed by semi-automatic annotation of video images. JAAD (Joint Attention in Autonomous Driving) is a popular behavior annotated dataset in this regard. Our contribution is to provide a knowledge graph based proposal of additional features that are important but missing in JAAD dataset in the context of road crossing decisions by pedestrians.