Road scene map for autonomous driving and modeling method

Constructing maps suitable for autonomous vehicles (AVs) is a critical research focus in autonomous driving and AI, extending cartography’s challenges. Building on cartographic principles, we propose the concept of a road scene map along with its modeling method that incorporates dynamic/static traffic elements with geometric/semantic features. Current limitations include unclear road scene graph relationships and a lack of integration among 3D traffic entity detection, map element detection, and scene relation extraction. To address these issues, we propose a method for constructing road scene maps: (1) A multi-task detection model identifies traffic entities and map elements directly in bird’s-eye-view (BEV) space, providing precise location, geometry, and attribute data; (2) A unified road scene relation pattern enables rule-based spatial/semantic relationship extraction. Experiments on nuScenes demonstrate improvements: the detection model achieves 1.5% and 1.9% accuracy gains in traffic entity and map element detection over state-of-the-art methods, while the relation extraction method covers broader perceptual ranges and more complex interactions. Results confirm the effective integration of 3D object detection, map element recognition, and scene relation extraction into a unified map. This integration delivers critical environmental information (locations, geometries, attributes, and spatial/semantic relationships) to AVs, significantly enhancing their perception and reasoning in dynamic road scenarios.