Accurate and Robust RGB-D Visual Odometry Based on Point and Line Features

Abstract

Feature recognition is widely used in visual simultaneous localization and mapping and visual odometry system. The most popular feature is the point feature, which includes SIFT feature, SURF feature, ORB feature. But in low textured scenes, RGB-D simultaneous localization and mapping (SLAM) tend to fail due to lack of reliable point features. Line features are as rich as point features in a structured or a low textured environment. However, line feature always contains more texture information for the calculation of pose, making it more complex for a line feature to be extracted, described and parameterized. To overcome the problems. We proposed a robust VO system fusing both points and lines by linearizing points, which not only preserves the information provided by line features, but also speeds up the calculation process. The method in this paper is evaluated on the TUM public real-world RGB-D dataset. The experimental results demonstrate that the algorithm proposed in this paper is more accurate and robust than the pure feature point algorithm in consideration.