Automatic vehicles are crucial components of intelligent transportation systems with significant applications in various fields. Road surface segmentation is the foundation for guiding autonomous vehicles to drive safely and achieve expected navigation. Although semantic segmentation is widely applied in autonomous driving, its performance in road surface segmentation is still limited in complex traffic environments. When given sparse known road information, this study proposes an adaptive prompt enhancement and fusion attention network for road surface segmentation. Specifically, we propose a multimodal feature embedding module to extract complementary features from multisource data, thereby improving the completeness of road surface modeling. Moreover, to address the issue of fragmented road surfaces, we design a progressive feature fusion module that adaptively integrates local road features with their global dependencies. Experiments demonstrate that the proposed model outperforms classical networks in segmentation performance, achieving a 13.22% improvement in IoU compared to using only single-source imagery. When given dense road information, we integrate trajectory density maps and prior road knowledge into the image channels as supplementary road features, which greatly enhances the ability of classical networks to extract roads. Experiments on two datasets demonstrate that the proposed model and strategies perform well in road segmentation accuracy, completeness, and topological connectivity.
Wang et al. (Thu,) studied this question.