Abstract: To improve the edge sharpness and small object recognition rate of street scene semantic segmentation, this paper proposes a street scene real-time semantic segmentation network(CitySeNet). The network was based on BiseNet V2, enhanced the detail branch with cross convolution blocks, optimized the semantic branch with cross-cross attention mechanism, and effectively fused the features of the two branches. The experimental results on the Cityscapes dataset show that the mean intersection over union(mIoU) of CitySeNet network is 77%, and the intersection over union of pedestrians and bicycles are 82% and 77% respectively. The visualization results of semantic segmentation show that the network can segment target edges more clearly, effectively solve the problems of edge blurring and inaccurate small object recognition, and meet the requirements of real-time segmentation.