Snake-SLAM: Efficient Global Visual Inertial SLAM using Decoupled Nonlinear Optimization

Abstract

Snake-SLAM is a scalable visual inertial SLAM system for autonomous navigation in low-power aerial devices. The tracking front-end features map reuse, loop closing, relocalization, and supports monocular, stereo, and RGBD input. The keyframes are reduced by a graph-based simplification approach and further refined using a novel deferred mapping stage to ensure a sparse yet accurate global map. The optimization back-end decouples IMU state estimation from visual bundle adjustment and solves them separately in two simplified sub problems. This greatly reduces computational complexity and allows Snake-SLAM to use a larger local window size than existing SLAM methods. Our system implements a novel multistage VI initialization scheme, which uses gyroscope data to detect visual outliers and recovers metric velocity, gravity, and scale. We evaluate Snake-SLAM on the EuRoC dataset and show that it outperforms all other approaches in efficiency while also achieving state-of-the-art tracking accuracy.