Crack segmentation is crucial in civil engineering, particularly for assessing pavement integrity and ensuring the durability of infrastructure. While deep learning has advanced RGB-based segmentation, performance degrades under adverse conditions like low illumination or motion blur. Thermal imaging offers complementary information by capturing emitted radiation, improving crack detection in challenging environments. Combining RGB and thermal images (RGB-T) for crack segmentation shows promise in complex real-world conditions, such as adverse weather, yet research in this area remains limited. Current RGB-T segmentation methods often fail to fully exploit the complementary relationships between modalities at various levels of interaction. To address this, we propose IRFusionFormer, a novel model for crack segmentation that effectively integrates RGB and thermal data. Our Efficient RGB-T Cross Fusion Module captures multi-scale relationships and long-range dependencies between modalities without significant computational overhead. Additionally, we introduce the Interaction-Hybrid-Branch-Supervision framework, which enhances interaction between modalities by distributing fused features across branches with joint supervision. To maintain the topological structure of cracks, we introduce a novel topology-based loss function that preserves connectivity during training. Our method achieves state-of-the-art performance, with a Dice score of 90.01% and an IoU of 81.83%, significantly improving robustness and accuracy in varying environmental conditions. These advancements address key challenges in pavement crack segmentation, offering a more reliable and efficient solution. For access to the codes, data, and models from this study, visit https://github.com/sheauhuu/IRFusionFormer