Robust RGB-D visual odometry with adaptive selection of point, line, and plane features

To deal with low-texture scenes, many visual simultaneous localization and mapping (SLAM) methods have introduced line features and plane features to provide additional structural information for stable frame tracking. These methods typically use a certain type of feature or feature combination, such as point-line features or point-line-plane features, throughout the entire camera pose estimation process. However simply increasing the number of feature types will introduce more noise sources, reducing the accuracy of camera pose estimation, thereby decreasing the localization accuracy. To solve this problem, this paper proposes an RGB-D visual odometry method that can automatically select different feature modes. It adaptively selects the type of features for frame-to-frame tracking based on the numbers of features extracted from current frames. The proposed method is evaluated on the TUM-RGBD dataset and achieves better trajectory accuracy than other algorithms.