Unsupervised domain adaptation technology is key to reducing the need for data labeling in computer vision tasks and implementing intelligent perception in equipment. Faced with the dispersion of feature distribution and class imbalance in real scenes (i.e., the target domain), such as blurry class boundaries and scarce samples, this paper proposes a Prototypes-Guided Adaptive Feature Fusion Model. It incorporates a Prototype-Guided Dual Attention Network that blends spatial and channel attention features to enhance class compactness. Moreover, an adaptive feature fusion module is introduced to flexibly adjust the importance of each feature, enabling the model to capture more class-discriminative features across different spatial locations and channels, thereby further improving semantic segmentation performance. Experiments on two challenging synthetic-to-real benchmarks, GTA5-to-Cityscape and SYNTHIA-to-Cityscape, validate the effectiveness of our method, demonstrating its advantages in dealing with complex scenes and data imbalance issues, and providing robust support for the visual perception technology of intelligent equipment.

domain adaptation, semantic segmentation, intelligent sensing, attention mechanism, self-training learning