Combined 2D and 3D features for robust RGB-D visual odometry

Abstract

As a novel type of sensor, RGB-D cameras have attracted substantial research at tention in indoor SLAM because they can provide both RGB and depth informa tion. Currently, most existing mature RGB-D SLAM solutions are keypoint-based, which su↵er from significant performance degradation in textureless scenes due to the lack of keypoints. Some works attempt to address this issue by incorporating line features. However, these methods still extract line features merely based on 2D RGB images, resulting in a restricted utilization of the environment’s 3D structural information and therefore providing only limited performance improvement. This project focuses on the fusion of 2D and 3D features for a robust RGB-D SLAM system. The proposed visual odometry extracts point, line, and surface features in the front-end to fully utilize the environment’s texture and structural information. In the back-end, a combination of loosely-coupled and tightly-coupled schemes is designed for multiple features to ensure both robustness and scalability of the system. Compared to existing state-of-the-art RGB-D SLAM systems, the e↵ectiveness and robustness of the proposed method is verified by experimental results. The proposed approach performs well both in scenes with limited texture or illumination variations and common scenes.