부산대학교 도서관

Autonomous robots sometimes need to operate in low-light environments, which can present significant challenges for simultaneous localization and mapping (SLAM) technology. Traditional SLAM systems based on visible-light cameras struggle to function effectively under low-light conditions, leading researchers to explore alternative SLAM systems based on thermal sensors. However, thermal sensors have limited feature information due to their low resolution. To fully leverage the information from both visible and thermal images while minimizing hardware dependency, this study proposes binocular visible-thermal (BVT)-SLAM, which is based on the combination of a monocular visible camera and a monocular thermal sensor. By utilizing multispectral stereo matching and feature point combination, the BVT-SLAM system takes advantage of complementary and spatial information between visible and thermal images, and can adaptively switch between monocular and stereo modes as needed. Nevertheless, due to the inability of traditional multispectral matching methods to meet real-time requirements, this study proposes an efficient feature point matching method based on bi-directional BRIEF descriptors (BDBRIEFs) that enables the system to operate in real-time. Experiments on multispectral datasets in different conditions demonstrate that the BVT-SLAM system is more robust in low-light scenes and provides higher positioning accuracy and real-time performance.