Visual-Tiered LiDAR-Inertial-Visual Odometry

This paper proposes a quality-guided LiDAR-Inertial-Visual odometry. Visual feature matching frequently suffers from mismatches and interference from dynamic objects. Existing visual odometry methods typically discard low-quality visual information, which can lead to insufficient information and reduced diversity. In visual-assisted odometry systems, high-quality visual features improve localization accuracy, while the impact of low-quality information on odometry performance depends on its distribution. This paper introduces a quality-filtered multi-hypothesis visual-inertial localization method, which retains visual features of varying quality levels and evaluates the reliability of these multi-hypothesis visual-inertial odometry models using laser localization results, enabling adaptive mode selection. By modeling the information from various sensors as error factors and considering sensor failures, we develop a factor-graph based LiDAR-Inertial-Visual odometry. The proposed method’s performance is validated using public datasets and our own constructed dataset. Experimental results demonstrate that, compared to the LVI-SAM algorithm, the proposed method achieves superior odometry accuracy.