An End-to-End Lane Detection Framework Based on Geometry Transform

Abstract

3D lane line detection plays an important role in Lane-keeping System, Lane-centering Assist, for intelligent vehicles. Most vision-based methods estimating 3D coordinates of lane lines rely on the inverse-perspective transformation, which affected by the road condition. However, This paper proposes a novel lane line detection framework that is immune to changes in terrain. The proposed framework includes an encoder and two decoders. First, image features are extracted by the feature encoder. Then, the duel-decoder architecture ensures the integrity of the semantic information of the lane lines in the initial image. The correlation between the lane lines is generated by the attention mechanism. Finally, the depth decoder’s output is combined through the geometry transform to obtain the 3D lane line directly. The proposed method explicitly handles the lane line occlusion. Experiments show that our framework has good performance in different driving scenarios.