Multi-Class Lane Semantic Segmentation of Expressway Dataset Based on Aerial View.

Multi-Class Lane Semantic Segmentation (MCLSS) is a hot topic in the computer vision research, which is of great significance for detecting violations of vehicles on expressway. At present, there is a lack of public dataset for semantic division of complete road areas. This paper completes the collection, cleaning, analysis, classification and labeling of expressway data on aerial view and proposes a semantic segmentation model "Deeplab-ERFC" (DeepLab with Erosion Loss and a Fully-Connected Conditional Random Field). This model extends DeepLabv3+ by adding a Erosion Loss (ER Loss) that can improve boundary prediction performance using corrosion operation to estimateHausdorff Distance (HD) and a Fully-Connected Conditional Random Field (Fully-Connected CRF) that will reduce the generation of cavities through two gaussian kernel functions considering the color intensity and position relationship between pixels. Based on mean Intersection over Union (mIoU), our proposed Deeplab-ERFC achieves the best semantic segmentation performance on our Expressway Dataset, reaching 83.9% mIoU in the test set.