To address the problems of difficult street object recognition and slow segmentation due to low visibility in foggy scence, a foggy cityscapes semantic segmentation algorithm incorporating self-supervised contrastive learning is proposed. The algorithm selects the lightweight network MobileNetV2 as the backbone network. Deep aggregation atrous spatial pyramid pooling module is designed and a deep separable convolution with dilation rate is used to replace the normal convolution to enrich feature diversity. Then, we increase the similarity of semantically similar pixels and maintain the distance between different semantic pixels by fusing the contrastive learning framework, so as to improve the model’s ability to represent and discriminate the detailed edges of small target objects. Finally, a new fusion loss function is proposed, and supervised learning and self-supervised learning are used to jointly guide the network training to learn deep feature representation. The experimental results show that the model can achieve MIoU of 74.35%, MPA of 83.59%, and PA of 95.85% on the Foggy Cityscapes dataset, which improves 3.82%, 3.99% and 1.02% respectively, compared with the semantic segmentation network DeepLabV3+ model. Meanwhile, the number of model parameters is 2.88M, which is nearly 55% less than the number of DeepLabV3+ model, optimizing the network computation consumption. The algorithm has good performance in foggy cityscapes semantic segmentation, reducing the number of model parameters while maintaining high segmentation accuracy and good robustness.