Sparse representation models perception as a set of discrete elements such as objects and lane lines. In crowded or occluded scenes this formulation struggles with unstructured obstacles, uncertain regions and intricate interactions. This paper introduces a dense risk‑aware occupancy representation that captures planning‑relevant risks in an explicit and uniform way. The representation builds a unified bird‑eye‑view (BEV) map that jointly encodes global scene occupancy, map‑derived traffic constraints and predicted future occupancy of dynamic agents, thus providing risk evidence across spatial and temporal dimensions. Based on this, we propose an end‑to‑end network called ROIDrive. An independent branch predicts risk‑aware occupancy, which is then injected into planning queries to produce safety‑oriented trajectories. To evaluate safety quantitatively we construct the RiskOcc4D‑nuScenes dataset on top of nuScenes and occ3d‑nuscenes. Results show that the risk‑aware occupancy reduces open‑loop collision rates by 52.9% under the UniAD metric and by 35.0% under the ST‑P3 metric on nuScenes.
Review