<p>Semantic segmentation provides technical support for automatic road sensing, and widely used in autonomous driving. However, when faced complex road scenes, existing segmentation methods still have problems such as insufficient learning of contextual information, inability to accurately segment multi-scale targets, and inaccurate segmentation of edge regions. To address these problems, we developed a multi-branch semantic segmentation method with CNN and Transformer (MBFormer). First, an exponentially weighted strip-window fusion pooling method in the CNN feature extraction stage was used to enhance the contextually important feature representation, enhance the long-distance dependencies, and expand the sensory field through a long and narrow kernel aggregation method. Then, cross-resolution multi-scale attention was used in the transformer feature extraction stage to fuse high and low resolution information to fully capture multi-scale information and enhance the segmentation capability of multi-scale targets. Finally, to solve the problem of inaccurate target edge segmentation, a branch based on boundary prediction was designed, and a boundary-aware module was used to establish the mapping relationship between pixels at different locations to fully capture the boundary information. Experiments were conducted on the Cityscapes and ACDC road datasets, and the mIoU of MBFormer achieved nearly 2% improvement compared to current mainstream segmentation models.</p>