Position Encoding for 3D Lane Detection via Perspective Transformer

3D lane detection from the input monocular image is a basic but indispensable task in the environment perception of automatic driving. Recent work uses modules such as depth estimation, coordinate system transformation, and time series tracking to achieve the correspondence of 2D to 3D information. However, inaccurate depth information generated by perturbations during conversion poses a challenge to lane detection methods that rely only on monocular images. To solve the above problems, we propose a PELD model, a proxy transformation using BEV aerial view, to explicitly give 3D lane detection results. Specifically, when sampling feature information, feature flipping is proposed to supplement the global context information before view conversion, and the 3D position encoding information generated by the forward-looking features enhances the depth information. After the 3D position encoding information is combined with the feature information, the cross-attention module is used as a value for adaptive supervision of BEV queries. On the one hand, we use deformable attention to sample forward looking features and generate explicit lane representation; on the other hand, we supplement supervised lane line generation by supplementing forward looking features and enhancing 3D spatial information. PELD implements a more advanced approach than ever before on OpenLane and Apollo datasets.