PLS-VIO: Stereo Vision-inertial Odometry Based on Point and Line Features
Huanyu Wen, Jindong Tian, Dong Li
Abstract
Traditional point feature-based vision SLAMs are difficult to find reliable point features to estimate camera pose in a weakly textured long corridor environment. Visual only SLAMs easily lose point features in weak texture or fast motion situations, causing the system to crash. And monocular VIOs have unobservable IMU scales under uniform motion. In view of the above, we propose PLS-VIO, an optimized Stereo vision-inertial odometry system using tight fusion of point and line features in this paper. PLS- VIO is divided into two threads: front-end point-line features tracking and back-end pose optimization. We use Plücker coordinates and orthonormal representation to represent line features, and optimize the 6-D pose by minimizing the objective function (including IMU error, reprojection error of point and line features, and prior error). We also improved the matching filtering strategy for line features. Our method is validated on the currently most popular public dataset (EuRoC), and shows superiority over the state of arts such as VINS-Mono, VINS-Fusion, OKVIS, MSCKF-Line, PL-VIO (no loop detection) etc. In the actual low texture and long corridor environment, the accuracy is improved by more than 50% compared with point feature-based methods.