Visual-inertial odometry with adaptive image enhancement for low-light

Abstract

Accurate state estimation is fundamental to the autonomous navigation of mobile robots. However, traditional visual-inertial odometry (VIO) often suffers from severe performance degradation in environments with insufficient illumination and sparse textures. To address this challenge, this paper proposes a robust, tightly-coupled monocular VIO algorithm leveraging adaptive image enhancement. In the frontend processing, an improved adaptive gamma correction is proposed for image enhancement to restore feature saliency under adverse lighting conditions. Furthermore, a cascaded outlier rejection scheme is designed, which incorporates a local flow consistency method as a pre-filtering step before the traditional random sample consensus, ensuring reliable tracking while significantly reducing computational overhead. The backend employs a sliding-window-based nonlinear optimization framework to achieve deep fusion of visual observations and inertial measurements within a probabilistic model. We evaluated the proposed algorithm on the public European robotics challenges micro aerial vehicle dataset and in challenging real-world scenarios. Quantitative results demonstrate that the proposed method outperforms VINS-Mono by an average of 24.7% in terms of positioning accuracy on the EuRoC dataset. In practical low-light tests, the accuracy gain reaches 51.0% compared to VINS-Mono. The results indicate that the proposed VIO system exhibits superior robustness and higher localization precision under illumination-degraded conditions.