ASLIF: Accurate Lane-Area Perception Through Adaptive Fusion of Structural Lines and Instance Segmentation

Lane-level road inventory aims to jointly extract direction-consistent lane areas and visible lane counts from street-view images. However, existing line-based methods do not directly recover complete lane regions, whereas area-based methods may include irrelevant pavement. To address these challenges, we propose adaptive structural line-instance fusion (ASLIF), a prediction-level framework that combines structural line predictions with instance-level area predictions. ASLIF converts O2SFormer line predictions into regional masks and fuses them with the continuous regional response produced by Mask R-CNN. An image-level fusion coefficient adjusts the relative contributions of the two branches according to scene complexity, estimated lane density, and the difference in branch confidence. Experiments on a test set from Nanjing, China, show that ASLIF achieves a precision of 0.92, a micro-F1 of 0.90, a mean-F1 of 0.90, and a mean-IoU of 0.82, with the highest precision among the evaluated baseline semantic-segmentation methods. These results indicate that ASLIF provides a precision-oriented operating point that suppresses false-positive lane regions while retaining substantial regional coverage.