Enhanced lane detection for autonomous driving based on ENet with attention-refinement

Abstract

Efficient lane detection is essential for autonomous driving because lane markings are thin, sparse, and vulnerable to shadow, occlusion, road wear, and background road markings. This paper presents an ENet-based lane segmentation method enhanced by a compact attention-refinement block. The block is explicitly defined as a combined channel, directional, and spatial gating mechanism, and a block diagram is added to clarify the computation of attention weights. Channel gating emphasizes lane-sensitive feature channels, directional gating preserves long lane continuity, and spatial gating suppresses isolated background responses. A half-resolution refinement path further fuses shallow edge information into the decoder. Five-fold validation on an extracted TuSimple subset gives an F1 score of 0.5524 and an IoU of 0.3835, improving over the ENet baseline by 0.0153 F1 and 0.0144 IoU. A small external CULane subset is also added as a stress test; the result shows a clear domain gap and motivates full CULane evaluation. On an NVIDIA GeForce RTX 4060 Laptop GPU, the model has 0.2740 M parameters, 5.0950 G FLOPs, and 230.46 FPS at 512 by 288 input resolution.