Graph-based topology reasoning for driving scenes
Tianyu Li, Li Chen, Huijie Wang, Yang Li, Jiazhi Yang, Xiangwei Geng, Shengyin Jiang, Yuting Wang, Hang Xu, Chunjing Xu, Junchi Yan
Abstract
Abstract Understanding the road structure is essential for achieving autonomous driving. This intricate topic contains two fundamental components: the interconnections between lanes and the associations between lanes and traffic elements (e.g., traffic lights), where a comprehensive topology reasoning method is still absent. On one hand, existing map learning techniques face challenges in deriving lane connectivity using segmentation or laneline-based representations; or prior approaches focus on centerline detection while neglecting interaction modeling. On the other hand, the topic of assigning traffic elements to lanes is limited in the image domain, leaving the construction of the correspondence between image and 3D views an unexplored challenge. To address these issues, we present TopoNet, an end-to-end topology reasoning network for analyzing driving scenes. To capture the topology of driving environments effectively, we introduce three key designs: (1) an embedding module that integrates semantic knowledge from 2D elements into a unified feature space; (2) a curated scene graph neural network that models relationships and facilitates feature interactions within the network; (3) a scene knowledge graph devised to differentiate prior knowledge from various types of the scene topology avoiding arbitrary message transmission. We evaluate TopoNet on the challenging scene understanding benchmark, OpenLane-V2, where our approach outperforms all previous studies by a significant margin across all perceptual and topological metrics. The code is released at https://github.com/OpenDriveLab/TopoNet .