An Illumination-Robust Real-Time Hybrid Visualinertial SLAM Framework Based on Deep Learning Networks

The paper seeks to improve the efficacy of visualinertial fusion SLAM under conditions characterized by fluctuating illumination, swift motion, and complex environments. A novel iterative technique for image enhancement is proposed. This algorithm mitigates brightness flickering between adjacent frames by utilizing the consistency of brightness across consecutive image intervals, which arises from independent calculations of single frames. This procedure creates advantageous conditions for further feature extraction. A multi-stage progressive adaptive hybrid weighting approach is developed to amalgamate filtering and optimization algorithms. This methodology maximally leverages the synergistic benefits of both techniques, integrating the advantages of Robocentric filtering and Worldcentric optimization algorithms while mitigating their individual limitations. Experiments utilize the Euroc dataset and the TUMVI dataset. The findings indicate that the suggested technique exceeds analogous methods in localization accuracy, demonstrating enhanced performance.