HSR-VINS: Hybrid Structural Regularities for Robust Visual-Inertial Navigation System

Visual-inertial navigation systems (VINS) are critical for aerial vehicles and robotics. However, traditional sparse point-based VINS are limited by the number of tracked points and do not effectively utilize structural information in the environment. To address this, we propose a visual-inertial navigation system with hybrid structural regularities, named HSR-VINS. This system integrates camera and inertial measurement unit (IMU) data and leverages multiple structural features in a non-independent manner. In the front-end, HSR-VINS introduces plane feature extraction and merging methods, ensuring robust detection and tracking of planes. It also develops a direction-guided merging approach to efficiently generate reliable line features and extracts vanishing points using a 2-line exhaustive search to capture the structural regularities. Additionally, the co-planarity of lines is determined using a 2D convex hull-based method to avoid feature redundancy. In the back-end, multiple structural regularities are integrated into the pose estimation process through residuals including line re-projection, vanishing points, point-plane, and line-plane residuals. Our method achieves an average RMSE APE of 0.07 m on the EuRoC dataset and 3.41 m on the KITTI dataset.