<p>Recognizing urban buildings at night is crucial for autonomous vehicles to navigate urban environments. However, under light-polluted nighttime conditions, the accuracy of semantic segmentation tends to degrade, impairing vehicle perception and representation learning. To address this challenge, we exploit the cross-domain stability of building depth and introduce depth as a geometric constraint by incorporating a building-aware geometric bias into the self-attention module of an unsupervised semantic segmentation framework. This depth-based geometric biasing approach offers a lightweight alternative to dual-encoding methods, which separately process depth maps and RGB images. Additionally, the original loss function is enhanced with a gradient-weighted loss tailored specifically for building structures. Experiments conducted in the complex urban road environments of Nanjing demonstrate that the proposed depth-assisted unsupervised segmentation method consistently outperforms baseline approaches for building segmentation in light-polluted nighttime conditions.</p>