EPS-Net: lane detection method based on EfficientNet and pyramid split attention

Lane detection is a cornerstone technology for autonomous driving. To improve both feature representation and geometric consistency under real-time constraints, we introduce three upgrades to the UFLD grid-classification framework. First, the ResNet backbone is replaced with EfficientNet-B2—leveraging MBConv blocks and squeeze-and-excitation attention—to deliver lightweight, high-capacity multi-scale features and faster inference. Second, a Pyramid Split Attention (PSA) module is incorporated to strengthen spatial–channel discrimination. Third, we devise a robust shape loss that couples the Huber criterion with lane-edge weighting to better enforce geometric coherence. Compared with the original UFLD, the proposed approach raises accuracy on TuSimple from 95.81% to 96.07% and boosts the F1 score on CULane from 68.4% to 69.5%, achieving notable accuracy gains while maintaining real-time performance.