End-to-end autonomous driving based on image plane waypoint prediction

Mainstream autonomous driving algorithms rely on high-definition (HD) maps, but HD maps are time-consuming and involve tedious manual labor. Some studies have adopted a simple end-to-end approach for autonomous driving, where inputed sensor informations are directly mapped to vehicle control quantities. This approach does not rely on HD maps, but is closely related to vehicle dynamics and is difficult to transfer directly between different vehicles, while being less capable of interpretation and generalization. To address the above issues, this paper proposed an end-to-end autonomous driving algorithm based on image-plane waypoints prediction, where the waypoints is an intermediate state quantity from input end to control end. Our method does not rely on HD maps, but also has good interpretability and generalization.