AFSAF-Net: Adaptive Frequency-Spatial Attention Fusion Network with IMU Guidance for Robust Indoor Localization under Low-Light and Texture-Degraded Environments

Abstract Indoor localization under low-light and texture-degraded environments remains challenging because insufficient visual information leads to unstable feature extraction, while inertial measurements are inevitably affected by accumulated drift. Previous MC4L-IMU-based studies demonstrated the effectiveness of handcrafted texture descriptors, frequency-domain enhancement, deep visual representation learning, and visual-inertial fusion for improving localization performance in dark indoor scenarios. However, these approaches mainly rely on manually designed frequency representations and fixed fusion strategies, limiting their adaptability to dynamically changing sensor reliability. To address these limitations, this paper proposes an Adaptive Frequency-Spatial Attention Fusion Network with IMU Guidance (AFSAF-Net) for robust indoor localization. The proposed framework introduces an adaptive frequency learning branch to capture spectral information beyond handcrafted transformations, an attention-driven frequency-spatial fusion module to dynamically integrate complementary visual representations, and an IMU reliability gate to adjust the contribution of inertial information according to sensor confidence. By jointly modeling frequency characteristics, spatial semantics, and inertial reliability, AFSAF-Net enables adaptive multimodal perception under challenging indoor conditions. Experiments conducted on the MC4L-IMU indoor localization platform demonstrate that AFSAF-Net achieves a mean localization error of 0.85 cm and the highest coefficient of determination (R 2 = 0.9986) among the evaluated methods. Furthermore, the proposed framework maintains more stable localization performance under low-light conditions compared with conventional visual and visual-inertial approaches. These results indicate that adaptive spectral representation learning and reliability-aware sensor fusion can effectively improve localization robustness without relying on fixed feature extraction or predefined fusion rules.