Transformer Generates Conditional Convolution Kernels for End-to-End Lane Detection

Environmental perception is crucial in autonomous driving technology, providing essential prior information for vehicle control and decision-making. Lane detection is gaining increasing attention as a critical component of environmental perception tasks. Mainstream frameworks rely on CNN-based architectures that often require preprocessing and postprocessing, complicating the implementation of end-to-end detection. Although Transformer-based structures address this issue from a global perspective, achieving higher accuracy remains challenging. This article proposes a novel Transformer-based end-to-end architecture, CondFormer, that processes each lane line separately, enhancing both accuracy and speed. First, we design a new Transformer structure to generate lane-by-lane parameter matrices from a global perspective instead of extracting features, efficiently constructing per-lane conditional convolution kernels. Second, to fully utilize multiscale information, CondFormer performs conditional convolution on the fused feature data. Then, each lane is further processed to obtain lane detection results using the ROW-wise method. Extensive experiments on the CULane, TuSimple, and CurveLanes datasets demonstrate that our method outperforms all Transformer-based end-to-end methods, offering a superior tradeoff between accuracy and speed. Our code is available at
https://github.com/Zhuanglong2/Condformer
.