Lane Detection by Variational Auto-Encoder With Normalizing Flow for Autonomous Driving

Abstract

Mainstream lane detection methods often lack flexibility, accuracy, and efficiency in challenging scenarios, especially with occlusion and extreme lighting. To address this, we reframe lane detection as a variational inference problem. Specifically, we propose a Variational Lane Detection Network (VLD-Net) using a Conditional Variational Auto-Encoder (CVAE) as the generative network to produce multiple lane maps as candidates, supervised by the ground-truth lane map. To build a more complex, expressive probability distribution, we incorporate normalizing flows into lane map generation, enhancing realism. Additionally, we develop a Lane-Attention Fusion (LAF) module using attention mechanisms to adaptively fuse generated candidate lane maps. LAF also includes a lane local feature aggregator to enhance local lane keypoint correlation. Experimental results on TuSimple and CULane datasets show our method outperforms previous approaches in challenging scenarios.