UPLP-SLAM: Unified point-line-plane feature fusion for RGB-D visual SLAM

Abstract

Most of the existing RGB-D simultaneous localization and mapping (SLAM) systems are based on point features or point-line features or point-plane features. However, the existing multi- feature fusion SLAM methods based on the filter framework are not accurate and robust. And the fusion methods based on the optimization framework process different kinds of features separately and integrate them loosely. In the optimization-based framework, how to tightly fuse various kinds of features for achieving more accurate and robust pose estimation has been rarely considered. In this paper, we propose a unified point-line-plane fusion RGB-D visual SLAM method for navigation of mobile robots in structured environments, making full use of the information of three kinds of geometric features. Specifically, it extracts point, line, and plane features using images captured by the RGB-D camera and expresses them in a uniform way. Then, a mutual association scheme is designed for data association of point, line and plane features, which not only considers the correspondence of homogeneous features, i.e., point-point, line-line and plane-plane pairs, but also includes the association of heterogeneous features, i.e., point- line, point-plane and line-plane pairs. Afterwards, the matching errors between homogeneous features and the association errors between heterogeneous features are uniformly represented and jointly optimized to estimate the camera pose and feature parameters for accurate and consistent localization and map building. It is worth pointing out that the proposed unified framework contains two levels. From the system framework perspective, all the main components of the SLAM system, i.e., feature representation, feature association and error function are handled in a unified manner, which increases the accuracy and compactness of the multi-feature SLAM system. From the feature processing perspective, both homogeneous features and heterogeneous features are uniformly used, which provides more spatial constraints on pose estimation. Finally, the accuracy and robustness of the proposed method are verified by experiment comparisons with state-of-the-art visual SLAM systems on public datasets and in real-world environments