Semantic segmentation in autonomous driving requires both accuracy and real-time performance. Since semantic segmentation necessitates features from both spatial and semantic information, most algorithmic models to achieve a larger receptive field and improve segmentation accuracy, result in a significant increase in parameters and a notable decrease in inference speed. By leveraging the advantages of dual-branch networks, which allow for parallel computation while retaining both detail and semantic information, we optimized the semantic branch and feature fusion module. Using the public Cityscapes dataset, we validated the model's segmentation effect and inference speed, achieving varying degrees of improvement.