ABSTRACT Visual–inertial navigation systems (VINSs) are a cornerstone technology for autonomous robotics, leveraging visual measurements and inertial sensor data to achieve accurate state estimation through feature tracking. However, current VINS approaches often indiscriminately incorporate all feature tracks without distinguishing which ones meaningfully contribute to estimation quality. This oversight leads to suboptimal accuracy, reduced stability, and computational inefficiency. To address these limitations, we propose a novel visual–inertial framework that intelligently prioritizes informative feature tracks using normal epipolar geometry. Our method enhances state estimation by jointly optimizing feature appearance similarity and observation consistency across frames, formalized as a submodular partition optimization problem. Extensive experiments on public benchmarks demonstrate that our approach significantly outperforms five state‐of‐the‐art methods, achieving a 26.1% improvement in accuracy, 43.3% higher stability, and real‐time processing speeds of up to 125 FPS. These advancements highlight the efficacy of selective feature track utilization in overcoming the limitations of conventional VINS.
Zhao et al. (Mon,) studied this question.