<![CDATA[Nighttime semantic segmentation remains a critical challenge for autonomous driving perception system due to the domain shift, which leads to substantial performance degradation when models trained on daytime data are directly applied to low-light environments. This study investigates the effectiveness of unsupervised domain adaptation and analyses the trade-off between segmentation accuracy and computational efficiency in nighttime scenarios. This research adopts the One-Stage Domain Adaptation Network for Unsupervised Nighttime Semantic Segmentation (DANNet) with a PSPNet backbone, which integrates a relighting network and adversarial learning to mitigate domain shift. Experiments are conducted using Cityscapes at the source domain and a locally collected UGM dataset as the target domain. The result show that DANNet improves segmentation performance from 5.27% to 9.29% under the 19-class configuration and achieves 22.09% mIoU in the 7-class configuration. The 7-class setup also demonstrates better computational systems, achieving 17.90 ± 0.60 FPS and a latency 55.94 ± 1.91 ms, making it suitable for near real-time deployment, whereas the 19-class configuration remains relevant for applications requiring finer semantic granularity.]]>
Copyrights © 2026
