RGB-Thermal semantic segmentation in nighttime traffic scenes is significant for autonomous vehicles to perceive the surrounding environment. However, finely labeled RGB-Thermal datasets of traffic scenes are scarce due to the challenge of pixel-level manual semantic labeling at night. To tackle this issue, we propose UDA-KB, a novel framework for unsupervised domain adaptation in RGB-Thermal semantic segmentation, which transfers knowledge from large-scale labeled RGB domain to unlabeled RGB-Thermal domain, circumventing the necessity for manual labeling. UDA-KB consists of three phases: source domain learning, day-to-night image translation, and target domain adaptation. Initially, we pre-train a teacher network using a large-scale RGB dataset. Subsequently, we present CUT-T (Contrastive Unpaired Translation-Thermal), which translates daytime images into nighttime equivalents by integrating thermal and boundary constraints, mitigating artifacts and alignment discrepancies in translating images. Finally, we employ the teacher network to generate pseudo-labels for the RGB-Thermal images, which are used to train the student network alongside the translated images. Additionally, we propose a knowledge bridge for target domain adaptation that considers the inherent discrepancies between teacher and student networks, facilitating their information exchange. Qualitative and quantitative experiments on the Freiburg and MF datasets attest to the efficacy of UDA-KB, highlighting its effectiveness in nighttime semantic segmentation.