Nachricht

Am Mittwoch, den 6. Mai 2015, hält Anirudh Sarma um 17:30 Uhr im Raum P 1.4.17 einen Vortrag über seine Masterarbeit mit dem Titel:

Reconstruction of semantic floor plan representation from 3D point clouds

Abstract:

In order to comply with high-level tasks, oftentimes in cooperation with humans, it is imperative that autonomous mobile robots are able to actively perceive and subsequently interpret their environment. This can be achieved with the help of various sensors ranging from sonars, laser range finders to active light cameras that have become available during the last years. The GETbot is equipped with range sensors that are capable of acquiring panoramic 3D point clouds of the robot's environment. One of these 3D scans can consist of several million points reproducing the geometry of the environment. Due to their low level nature, high data quantities as well as sensor noise, point clouds are not the most suitable long-term representation. Especially for indoor environments that tend to offer highly planar surfaces it is reasonable to aim for a geometrical abstraction of particular surfaces within the point cloud. Human environments typically contain a range of different objects, such as chairs, tables, cabinets and shelves. Clutter, a random distribution of these types objects leads to confusion of surfaces, as well as occlusion effects. The latter causes that there are surfaces that are not acquired completely, leaving blind spots in the resulting point cloud. As soon as the reconstruction is applied to larger environments, such as hallways or lab rooms, the point density within the cloud will considerably vary and hence affect the reconstruction quality and success rate. In some cases, multiple scans need to be taken for one room, hence, partial clouds need to be registered before the reconstruction. This work aims to create a coherent 3D high-level reconstruction ranging over multiple rooms. Features representing transitions between rooms, commonly doorways, need to be extracted and matched, as the exact pose of each of the scans cannot be assumed.