Bai Yifeng ; Chen Zhirong ; Fu Zhangjie ; Peng Lang ; Liang Pengpeng ; Cheng Erkang

Abstract

3D lane detection is an integral part of autonomous driving systems. PreviousCNN and Transformer-based methods usually first generate a bird's-eye-view(BEV) feature map from the front view image, and then use a sub-network withBEV feature map as input to predict 3D lanes. Such approaches require anexplicit view transformation between BEV and front view, which itself is stilla challenging problem. In this paper, we propose CurveFormer, a single-stageTransformer-based method that directly calculates 3D lane parameters and cancircumvent the difficult view transformation step. Specifically, we formulate3D lane detection as a curve propagation problem by using curve queries. A 3Dlane query is represented by a dynamic and ordered anchor point set. In thisway, queries with curve representation in Transformer decoder iterativelyrefine the 3D lane detection results. Moreover, a curve cross-attention moduleis introduced to compute the similarities between curve queries and imagefeatures. Additionally, a context sampling module that can capture morerelative image features of a curve query is provided to further boost the 3Dlane detection performance. We evaluate our method for 3D lane detection onboth synthetic and real-world datasets, and the experimental results show thatour method achieves promising performance compared with the state-of-the-artapproaches. The effectiveness of each component is validated via ablationstudies as well.