<p>In urban planning, intelligent traffic management, and emergency response, accurate road information is of great importance. However, pixel-level annotation is costly, data are imbalanced, and road shapes vary widely. These issues lead to poor pseudo-label quality and insufficient detail representation in deep-learning-based road segmentation. To address these problems, we propose CrossViewSeg, a cross-view semi-supervised learning framework. First, a dual-network processes the same image under weak and strong augmentations, with perturbations introduced at both the image and feature levels. Then, through a cross-view consistency mechanism, pseudo-labels generated by the weak augmentation branch supervise the strongly augmented branch without thresholding. Next, a cross-attention feature enhancement module is introduced to maximize feature differences between the two paths and avoid homogenization. Finally, higher weights are assigned to regions of disagreement between the two views in the cross-entropy loss, so that learning focuses on hard-to-segment road networks and complex areas. Experiments on the DeepGlobe and Massachusetts datasets demonstrate that CrossViewSeg outperforms mainstream methods across multiple metrics, particularly showing stronger structural coherence and road boundary delineation in complex traffic scenarios.</p>