Abstract

Understanding the road genome is essential to realize autonomous driving.This highly intelligent problem contains two aspects - the connectionrelationship of lanes, and the assignment relationship between lanes andtraffic elements, where a comprehensive topology reasoning method is vacant. Onone hand, previous map learning techniques struggle in deriving laneconnectivity with segmentation or laneline paradigms; or prior lanetopology-oriented approaches focus on centerline detection and neglect theinteraction modeling. On the other hand, the traffic element to lane assignmentproblem is limited in the image domain, leaving how to construct thecorrespondence from two views an unexplored challenge. To address these issues,we present TopoNet, the first end-to-end framework capable of abstractingtraffic knowledge beyond conventional perception tasks. To capture the drivingscene topology, we introduce three key designs: (1) an embedding module toincorporate semantic knowledge from 2D elements into a unified feature space;(2) a curated scene graph neural network to model relationships and enablefeature interaction inside the network; (3) instead of transmitting messagesarbitrarily, a scene knowledge graph is devised to differentiate priorknowledge from various types of the road genome. We evaluate TopoNet on thechallenging scene understanding benchmark, OpenLane-V2, where our approachoutperforms all previous works by a great margin on all perceptual andtopological metrics. The code is released athttps://github.com/OpenDriveLab/TopoNet