Planar Features for Visual SLAM

Among the recent trends in real-time visual SLAM, there has been a move towards the construction of structure-rich maps. By using landmarks more descriptive than point features, such as line or surface segments, larger parts of the scene can be represented in a compact form. This minimises redundancy and might allow applications such as object detection and path planning. In this paper, we propose a probabilistic map representation for planar surface segments. They are measured directly using the image intensities of individual pixels in the camera images. Preliminary experiments indicate that the motion of a camera can be tracked more accurately than with traditional point features, even when using only a single planar feature.