STRUCTURAL LINE FEATURE SELECTION FOR IMPROVING INDOOR VISUAL SLAM

Nowadays, Visual SLAM has gained ample successes in various scenarios. For feature-based system, it is still limited when running in an indoor room, as the indoor scene is often with few and simple texture which result in less and unevenly distributed point features. To solve this limitation, line features which are quite rich in an indoor scene are extracted and used. However, not all features can geometrically contribute to pose estimation, specifically, line features that are consistent to the motion direction provide only weak geometric constraint for solving pose parameters. Therefore, this paper proposes a selection method for reasonable line features, in particular, based on the Manhattan World Assumption (MWA), structural line features are firstly extracted instead of normal line features. Then, the structural line features are selected according to the direction information of vanishing points and selected for a stronger geometric constraint on pose estimation. In general, the selected structural lines require that the intersection angle between the corresponding principal direction and the camera motion direction is higher than a threshold, which is extensively investigated in the experiments. The experimental results show that, compared to the original ORB-SLAM2, the localization accuracy after using the proposed method can be improved by around 15%-40% on various public datasets, and the real-time performance can be basically guaranteed even including the extra time spent on the selection procedure.