Improving SLAM on a TOF Camera by Exploiting Planar Surfaces

Abstract: Simultaneous localization and mapping (SLAM) is the problem of mapping your surroundings while simultaneously localizing yourself in the map. It is an important and active area of research for robotics. In this master thesis two approaches are attempted to reduce the drift which appears over time in SLAM algorithms. The first approach tries 3 different motion models for the camera. Two of the models exploit the a priori knowledge that the camera is mounted on a trolley. These two methods are shown to improve the results. The second approach attempts to reduce the drift by reducing noise in the point cloud data used for mapping. This is done by finding planar surfaces in the point clouds. Median filtering is used as an alternative to compare the result for noise reduction. The planes estimation approach is also shown to reduce the drift, while the median estimation makes it worse.