Abstract
Lane detection is a core perception task for lateral position estimation, lane keeping, and path planning in intelligent driving systems. However, in low-visibility environments such as rain, fog, and nighttime scenes, lane markings are affec...
Abstract
Lane detection is a core perception task for lateral position estimation, lane keeping, and path planning in intelligent driving systems. However, in low-visibility environments such as rain, fog, and nighttime scenes, lane markings are affected by low contrast, road-surface reflection, wiper occlusion, local discontinuity, and weak responses from distant lane markings. Previous studies have contributed to real-time detection and segmentation through the YOLO series, background suppression through important-region enhancement, shape adaptation through deformable convolution, and improved lane continuity through directional information propagation. Nevertheless, these approaches still have limitations in jointly explaining long-distance directional continuity and irregular shape variation of low-visibility lane markings within a progressive experimental framework.
To address these limitations, this study designs a spacial filter module using 3×3, 3×5, 5×3, 3×7, and 7×3 filters. By skipping one or two rows or columns during computation, the module expands the directional receptive field while maintaining a computational burden comparable to that of a 3×3 filter. Rather than simply stacking all modules, this study begins with the baseline YOLOv11 model, first verifies the effects of YOLOv11+EBAM and YOLOv11+DEC, then examines the extended directional receptive range provided by YOLOv11+spacial filter, and finally compares YOLOv11+EBAM+spacial filter with YOLOv11+DEC+spacial filter. The training data are reorganized from the public ETRI dataset with a resolution of 1920×1200 into the YOLO segmentation format, including 12,777 training images and 3,167 validation images. Among these, 91 images are rainy scenes. For separate real-environment validation, 50 rainy frames with a resolution of 1280×720, recorded directly by the researcher, are used.
The experimental results show that, in the training-stage comparison, YOLOv11+DEC achieves the highest Precision of 0.81. In the self-collected rainy frames, YOLOv11+DEC+spacial filter achieves the best performance, with an average IoU of 0.50 and an average Mask Precision of 0.53. Compared with the baseline YOLOv11 results of 0.36 and 0.32, these values correspond to improvements of approximately 38.89% and 65.63%, respectively. These results indicate that the improvement in low-visibility lane detection is mainly attributable to the combined effect of DEC-based shape adaptation and the spatial perception enhancement provided by the spacial filter.
Key words: low visibility; lane detection; YOLOv11; EBAM; deformable convolution; multi-scale directional convolution