IMPS: Informative Map Point Selection for Visual-Inertial SLAM

Visual-inertial SLAM (VINS) stands at the forefront of advancements in computer vision, robotics, and autonomous driving, revolutionizing the way we perceive and navigate the world. The typical approach in optimization-based methods for VINS is to estimate camera poses by minimizing the reprojection errors of all corresponding map points. However, not all map points are suitable for accurate pose estimation. Some map points introduce noise and can compromise the accuracy of the estimation. Current methods primarily focus on removing 2D noise points, which proves to be ineffective since triangulation can increase the uncertainty of map points. To address this problem, we propose a new method for Informative Map Point Selection (IMPS) in VINS systems. IMPS identifies the most informative map points by utilizing mutual and geometric information as the optimization target. We integrate IMPS into a VINS system and evaluate its performance using publicly available EuRoC, TUM and M2DGR datasets, as well as our own data. Experimental results demonstrate that our method surpasses existing VINS methods and achieves state-of-the-art pose estimation performance. Importantly, IMPS functions as an independent module with strong generalization capabilities, allowing for easy integration into other VINS systems and enhancing pose estimation performance. This integration not only elevates pose estimation accuracy but also holds great promise for advancing applications in intelligent driving and unmanned systems.