Structure-aware SLAM with planes and lines in man-made environment

Abstract This paper describes a robust approach to monocular SLAM by incorporating structural elements in man-made environments. Indoor and outdoor scenes often contain a large number of planes and lines, while the classic point-based methods tend to ignore the geometric clues in these common structures. The proposed method exploits not only strong structural constraints on points. More importantly, planes and lines are also treated as supplement for long-distance tracking. Drifting error is reduced via detecting large planes spanning over views and reinforcing coplanarity of triangulated points. Lines are integrated across views with an inverse depth representation, providing extra one-dimensional features and directional information. The structural constraints and additional features are incorporated under a unified minimization framework, and the planes and lines are actively extended as the SLAM system goes on, making our algorithm suitable for exploration-style applications. Real world monocular sequences have demonstrated that the proposed SLAM system outperforms the state-of-the-art and produces accurate results in both indoor and outdoor scenes.