Abstract:

Visual simultaneous localization and mapping (SLAM) is susceptible to tracking failure under complex illumination conditions such as low-light and high-exposure due to feature quality degradation. To address this challenge, we propose a visual SLAM method based on a lightweight illumination enhancement network. First, a spatially adaptive illumination equalization (IE) curve is designed to achieve low-light enhancement and high-exposure suppression through per-pixel spatial trend maps while maintaining photometric consistency in normally illuminated regions. Then, an image-mean-driven adaptive continuous iteration strategy is proposed to dynamically adjust the curve order for images with varying exposure levels, effectively avoiding the limitations of fixed iterations. Finally, we improve the deep curve estimation network (DCE-Net) by employing depth-wise separable convolutions (DSC) and end-batch normalization to maintain training stability with almost the same parameter count. Tests show that single-frame processing on the CPU requires only 12 ms, meeting the real-time requirements of the SLAM frontend. To validate the proposed method, experiments were conducted on 12 challenging illumination sequences from the EuRoC, TUM-VI, and KITTI datasets. Comparative analyses were performed with various traditional and deep learning visual enhancement algorithms. The experimental results demonstrate that the root mean square error (RMSE) of absolute trajectory error (ATE) is reduced by 32.3% on average compared with VINS-Fusion, achieving the best performance among all compared methods. This method provides a solution for the reliable deployment of visual SLAM in complex illumination environments.

Key words: visual SLAM, image enhancement, complex illumination, deep learning