Pittner Maximilian ; Janai Joel ; Condurache Alexandru P.

Abstract

Monocular 3D lane detection has become a fundamental problem in the contextof autonomous driving, which comprises the tasks of finding the road surfaceand locating lane markings. One major challenge lies in a flexible but robustline representation capable of modeling complex lane structures, while stillavoiding unpredictable behavior. While previous methods rely on fullydata-driven approaches, we instead introduce a novel approach LaneCPP that usesa continuous 3D lane detection model leveraging physical prior knowledge aboutthe lane structure and road geometry. While our sophisticated lane model iscapable of modeling complex road structures, it also shows robust behaviorsince physical constraints are incorporated by means of a regularization schemethat can be analytically applied to our parametric representation. Moreover, weincorporate prior knowledge about the road geometry into the 3D feature spaceby modeling geometry-aware spatial features, guiding the network to learn aninternal road surface representation. In our experiments, we show the benefitsof our contributions and prove the meaningfulness of using priors to make 3Dlane detection more robust. The results show that LaneCPP achievesstate-of-the-art performance in terms of F-Score and geometric errors.