Identifying Unobserved Road Regions in Bird's-Eye View for Single-Vehicle Perception

Abstract

Visual Bird's-Eye View (BEV) perception is a foundational paradigm in autonomous driving, enabling top-down semantic-spatial representations from multi-view inputs. However, current BEV methods struggle with occlusions, often generating hallucinated predictions in regions that are unobserved by the ego-vehicle. This mismatch between predictions and physical reality poses critical risks in safety-sensitive scenarios. To address this, we introduce a novel task: unobserved road segmentation, focusing on explicitly identifying road regions occluded from the ego-vehicle's view. To tackle this task, we propose Polar Dual-Attention Cross-View Transformers (PDA-CVT), an efficient query-based framework that leverages polar-coordinate cross-view attention and local self-attention to improve occlusion reasoning efficiently. Additionally, an automated ground-truth generation process alleviates the need for manual annotation. Experimental results show that PDA-CVT outperforms existing methods in accurately segmenting unobservable road regions as well as conventional road layouts while maintaining efficiency. This work represents a critical step toward safer and more reliable autonomous driving through enhanced occlusion-aware perception. Code and data are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/yujinyua/PDA-CVT</uri>.