Traditional visual-inertial SLAM (Simultaneous Localization and Mapping) systems often struggle to maintain stability in dark environments or under feature-degraded conditions, where conventional feature extractors are unable to provide reliable and consistent visual features, leading to poor tracking performance and severe trajectory drift. To overcome these limitations, we propose XfeatVINS, a robust monocular visual-inertial SLAM system designed specifically for challenging scenarios such as darkness, smoke, and underground spaces. XfeatVINS integrates infrared imaging with XFeat, a deep-learning-based feature extraction network tailored for extreme environments. The system adopts a tightly coupled visual-inertial fusion framework that combines high-frequency inertial measurements with robust visual features extracted from infrared images, significantly enhancing localization accuracy and robustness.To validate the effectiveness of XfeatVINS, we conduct extensive experiments on real-world datasets collected under harsh lighting and degraded visual conditions. The results demonstrate that XfeatVINS achieves significantly lower absolute trajectory error (ATE) and reduced trajectory drift compared to the widely used VINS-Mono system. Moreover, it consistently delivers reliable pose estimation where traditional visual front-ends fail.These results highlight the capability of XfeatVINS as a practical and scalable solution for visual-inertial localization in degraded environments, offering strong potential for applications in autonomous exploration, emergency response, and subterranean robotics.