Monocular Visual-Inertial Odometry with Point and Line Features Using Improved Line Feature Extraction

Visual-inertial odometry that uses point features only has poor localization performance in the low-texture environments, or even fails. The introduction of line features can improve this situation, because line features have richer scene structure information and can appropriately improve the positioning accuracy. We propose a real-time optimized tightly coupled visual-inertial algorithm, which based on point-line features. It can achieve a good balance in the positioning accuracy and the real-time performance. The front-end part proposes an algorithm for adjusting image brightness based on image preprocessing to increase the number of image feature extractions in low-light environments. In order to reduce the wrong segmentation of long line features, the least squares method is used to combine short line segments with similar directions and distances, thereby reducing the difficulty of matching line features. Experiments on the EuRoc dataset show that, under the same experimental environment, our algorithm has higher stability and positioning accuracy compared with the mainstream VIO algorithm based only on point features. At the same time, compared with PL-VINS which uses point-line features, our algorithm improves the positioning accuracy by 6%-7%.