Wang Ruihao ; Qin Jian ; Li Kaiying ; Li Yaochen ; Cao Dong ; Xu Jintao

Abstract

3D lane detection which plays a crucial role in vehicle routing, has recentlybeen a rapidly developing topic in autonomous driving. Previous works strugglewith practicality due to their complicated spatial transformations andinflexible representations of 3D lanes. Faced with the issues, our workproposes an efficient and robust monocular 3D lane detection called BEV-LaneDetwith three main contributions. First, we introduce the Virtual Camera thatunifies the in/extrinsic parameters of cameras mounted on different vehicles toguarantee the consistency of the spatial relationship among cameras. It caneffectively promote the learning procedure due to the unified visual space. Wesecondly propose a simple but efficient 3D lane representation calledKey-Points Representation. This module is more suitable to represent thecomplicated and diverse 3D lane structures. At last, we present a light-weightand chip-friendly spatial transformation module named Spatial TransformationPyramid to transform multiscale front-view features into BEV features.Experimental results demonstrate that our work outperforms the state-of-the-artapproaches in terms of F-Score, being 10.6% higher on the OpenLane dataset and5.9% higher on the Apollo 3D synthetic dataset, with a speed of 185 FPS. Thesource code will released at https://github.com/gigo-team/bev_lane_det.