Improving Visual SLAM in Complex Nighttime Scenarios: A Robust Low-Light Solution

Traditional visual SLAM systems often degrade in nighttime scenarios due to challenges such as intense light pollution and localized illumination blind spots. To address these issues, this paper proposes an enhanced solution based on the VINS-MONO framework, with improvements in three key components: front-end visual odometry, data initialization, and back-end optimization. First, a brightness-adaptive feature management module is introduced in the front-end design. This module dynamically monitors global average brightness and local illumination variations to adjust feature selection and optical flow tracking parameters. By calculating dynamic thresholds based on neighborhood brightness, overexposed feature points are filtered out, while the feature extraction interval in non-overexposed regions is reduced to mitigate sparsity. Second, a motion-state-aware robust initialization method is proposed to resolve prolonged or failed initialization under low-light conditions. Finally, in the back-end optimization, global brightness information is utilized to dynamically adjust the sliding window size and adaptively balance the weights of IMU and visual residuals in the optimization equation. Experiments conducted across daytime and nighttime scenarios, using both handheld and drone platforms, demonstrate that the improved system achieves higher robustness and localization accuracy in complex nighttime scenarios. These results validate the effectiveness of the proposed approach.