mVLINS: A Multilevel Visual-LiDAR-Inertial Navigation System with Completely Decoupled Odometry and Adaptive Environmental Mapping

The core component of the navigation system is the process of multi-sensor fusion localization in a challenging environment with limited GNSS constraints. This process is designed to robustly and precisely estimate the system state and map the traversed area. In this paper, we propose a low-drift, highly resilient multilevel navigation system with completely decoupled odometry and adaptive environmental mapping. A novel method for visual-LiDAR-inertial frame-to-frame odometry is presented to leverage the complementary strengths of these sensors. This odometry approach involves decoupling the 6-degree-of-freedom (6-DoF) state to allocate each state component to an appropriate submodule for estimation. An adaptive environmental mapping module that aims to align the target frame with the local map is proposed to refine the rough odometry pose. This module is achieved through the utilization of an adaptive keyframe strategy and the feature consistency constraint. Enhance the matching of keyframes to the map by reducing the point-to-feature error between frame points and map features. Additionally, the state is further refined by minimizing the feature consistency error on lines in the adjacent corner map and the local map at each keyframe. Our proposed algorithm is validated using both public and self-collected datasets, demonstrating superior results compared to state-of-the-art algorithms.