Automatic reconstruction of polygonal room models from 3D point clouds

The implementation of context-bound tasks at a higher level of abstraction often demands a semantic representation of the environment and the enclosed objects. A fundamental step towards the creation of rich semantic models is the abstraction of information from the raw and oftentimes unordered sensor data. In this work, we introduce a novel method for the abstraction of indoor environments as polyhedral models of interior rooms yielding to 3D floor plans. 3D laser scans that are acquired from a mobile robot serve as inputs for our method. First, the point cloud is segmented and polygonized. The final model is extracted by relating individual segments and determining the best configuration for the given room. With the help of a room model, we can easily discriminate between points that are located within the room and those that are part of the room boundaries. This facilitates succeeding tasks such as recognition of objects and furniture, detection of doorways as well as place recognition.