Point–Line Visual–Inertial Odometry With Optimized Line Feature Processing

Visual-inertial odometry (VIO) utilizing point and line features has become prevalent for state estimation in structured environments. The robustness of line segment detection and the efficiency of line feature matching significantly impact pose estimation performance. However, prevailing gradient-based line detection methods are prone to noise, and the existing line feature matching methods suffer from inefficiency, limiting the practical deployment of point-line visual-inertial odometry systems. To address these challenges, we present a novel point-line visual-inertial odometry system (POL-VIO), featuring advancements in line detection, matching, and back-end optimization modules. The proposed approach introduces an edge-constraint deep line detector to fortify line completeness and resilience to noise. Subsequently, the optical flow of points on lines is harnessed to establish initial correspondences, incorporating geometric linearity constraints to enhance the efficiency and accuracy of line tracking. In the final phase, distance and angle parameters are utilized to represent line reprojection error, minimizing point-line measurement residuals alongside inertial measurement unit (IMU) measurements for precise estimation of pose and feature depth. Extensive experiments conducted on both the public EuRoC dataset and our collected corridor dataset demonstrate the effectiveness of the proposed system. The results show that the localization error of our method is 28.5% less than that of PL-VINS, while achieving more than twice the line-tracking speed. The codes and dataset are available at https://github.com/HanqianSi/POL-VIO.