Combine HD map information to enhance road topology reasoning by offline supervised training and on-line ensemble learning

Topology reasoning plays a crucial role in understanding complex driving scenarios and facilitating downstream planning, yet the process of perception is inevitably affected by weather, traffic obstacles and worn lane markings on road surface. Combine pre-produced high-precision maps, and other type of map information to the perception network can effectively enhance perception robustness, but this on-line fused information often requires a real-time connection to website servers. We are exploring the possibility to compress the information of offline maps into a network model and integrate it with the existing perception model. We designed a topology prediction module based on graph attention neural network and an information fusion module based on ensemble learning. The module, which was pre-trained on offline high-precision map data, when used online, inputs the structured road element information output by the existing perception module to output the road topology, and the output road topology is input to the ensemble learner together with the topology output by the existing perception model for information fusion. Our method proposes a paradigm for utilizing offline high-precision map information through offline supervised learning and online ensemble learning. Experimental results show that it achieves varying degrees of algorithm accuracy improvement for multiple topology prediction algorithms on the OpenLane-V2 dataset.