Autonomous navigation requires structured representation of the road network\nand instance-wise identification of the other traffic agents. Since the traffic\nscene is defined on the ground plane, this corresponds to scene understanding\nin the bird's-eye-view (BEV). However, the onboard cameras of autonomous cars\nare customarily mounted horizontally for a better view of the surrounding,\nmaking this task very challenging. In this work, we study the problem of\nextracting a directed graph representing the local road network in BEV\ncoordinates, from a single onboard camera image. Moreover, we show that the\nmethod can be extended to detect dynamic objects on the BEV plane. The\nsemantics, locations, and orientations of the detected objects together with\nthe road graph facilitates a comprehensive understanding of the scene. Such\nunderstanding becomes fundamental for the downstream tasks, such as path\nplanning and navigation. We validate our approach against powerful baselines\nand show that our network achieves superior performance. We also demonstrate\nthe effects of various design choices through ablation studies. Code:\nhttps://github.com/ybarancan/STSU\n