Real-time semantic segmentation is a key branch of image segmentation in computer vision. Its goal is to predict the corresponding class for each pixel in an input image while meeting real-time requirements. To address issues in semantic segmentation algorithms, such as chaotic region segmentation, inaccurate segmentation of small objects, and discontinuous edge segmentation, this paper adopts a two-branch network architecture: the detail branch and the semantic branch extract detail and semantic information, respectively. The detail branch primarily captures local features such as details and textures, but lacks semantic understanding. The semantic branch extracts high-level semantic information. This type of information typically requires a large receptive field to capture global contextual features in the image. To alleviate these issues, we use a Pixel-attention-guided fusion module to integrate information from the semantic branch in the first two stages of the detail branch to enhance its understanding of the global context. Furthermore, we add a three-branch context dynamic fusion module at the end of the semantic branch. This module captures multi-scale contextual information through three parallel convolutional branches. In the semantic branch, a short-term dense cascade module with efficient channel attention is used to focus on key channels through an efficient channel attention module. The experimental results show that our model achieves 78.4% mIoU with inference speed of 65 FPS on Cityscapes and 76.0% mIoU with inference speed of 61 FPS on CamVid. Compared with other methods, this model improves segmentation accuracy while maintaining inference speed, achieving a balance between speed and accuracy.