LN-SLAM: deep learning-based underground low-light loop closure SLAM

SLAM technology shows great promise for dense simultaneous localization and mapping (SLAM) but faces challenges such as poor image quality and weak loop closure effects under low-light conditions in coal mine tunnels. We propose LN-SLAM, a deep learning-based low-light enhancement loop detection SLAM system. This method introduces IL-SLAM, which is based on IAT image enhancement and loop optimization. We incorporate the LSKAttention large kernel attention mechanism and SSIM loss loss function to develop the LSS-IAT low-light image enhancement network, which captures broader contextual information while enhancing image detail. The proposed ParC_NetVLAD image matching algorithm improves robustness in large-scale scenarios by using local rather than global information.