FGNet: Robust lane detection for autonomous driving via frequency-guided feature enhancement

Abstract

Lane detection is a critical component in autonomous driving perception systems. Complex road scenarios featuring varying lane appearances, challenging lighting conditions, and vehicle occlusions pose significant challenges for accurate lane detection. To address these problems, we propose FGNet, a robust lane detection framework that enhances feature representation through frequency-domain analysis and adaptive global-local fusion. We first introduce a Wavelet-enhanced Feature Pyramid Network (WLFPN) that leverages discrete wavelet decomposition and directional convolutions to capture high-frequency geometric features critical for lane structure modeling. Subsequently, a Global-Aware Feature Refinement (GAFR) module is designed to overcome insufficient global context integration in existing anchor-based methods, enabling adaptive feature enhancement through spatially-aware attention and selective fusion mechanisms. Finally, a Dynamic Loss Harmonizer (DLH) employs momentum-based dynamic weight adjustment to optimize multi-loss learning, improving training stability and convergence. Extensive experiments demonstrate that FGNet achieves state-of-the-art performance with F1 scores of 80.64 % and 97.89 % on the challenging CULane and TuSimple datasets, respectively, outperforming existing methods.