Design and implementation of an indoor modeling method through crowdsensing

Abstract

While automatic modeling and mapping of outdoor environments is well-established, the indoor equivalent of automated generation of building floor plans poses a challenge. In fact, outdoor localization is commonly available and inexpensive through the existing satellite positioning systems, such as GPS and Galileo. However, these technologies are not applicable in indoor environments, since a direct line of sight to the satellites, orbiting the globes, is required. As a substitution, the technical literature comprises several proposals for the development of simultaneous indoor localization and mapping (SLAM). In these approaches, the authors mostly exploit indoor resources such as the WiFi access points and the mobile smart devices carried by individuals in the indoor environment. Collecting data from several mobile devices is referred to as crowdsensing. To enable the generation of two-dimensional (2D) as well as three-dimensional (3D) maps, we propose crowdsensing of point clouds, which are 3D data structures of points in space. For localization, we integrate two features of a recently developed mobile device, called Project Tango. Specifically, the Tango platform provides two main technologies for reliable localization, namely motion tracking and area learning. Moreover, Tango-powered devices provide us with the ability to collect point clouds though a third technology, called depth perception. In the past few years, spatial data obtained from range imaging was used to generate indoor maps. Nevertheless, range images are expensive and not always available. The required equipment, e.g. laser range scanners, are both expensive in procurement and require trained personnel for proper setup and operation. In this thesis, we aim for obtaining spatial point clouds via crowdsensing. The main idea is to use sensor data which can be scanned by volunteering individuals using easy to handle mobile devices. Specifically, we depend on depth perception capabilities as provided by Google Tango-powered tablet computers. A crowdsensing infrastructure assigns scanning tasks to individuals carrying a Tango device. Execution of such a task consists of taking scans of e.g. offices in a public building. The scanning results contain both spatial information about the room layout and its position. Energy consumption on the mobile device is reduced by applying Octree compression to the scanned point clouds, which results in a significant reduction of the amount of data, which has to be transferred to a back-end server. Afterwards, the back-end is responsible for assembling the received scans and the extraction of an indoors model. The modeling process - developed in this thesis - comprises two-phases. First, we extract a basic model from the obtained point clouds, which may contain outliers, inaccuracies and gaps. In the second phase, we refine the model by exploiting formal grammars. It is worth to mention here that we are the first to exploit formal grammars as a model fitting tool. We feed the information obtained in the first phase to an indoors grammar, which has been developed in the ComNSense project, University of Stuttgart. The resultant model both contains much less deviations from the ground truth and provides improved robustness against aberrations with respect to localization during the scanning process. Thus, instead of scanning multiple point clouds per room, we need only one scan to be able to construct an indoor map. During evaluation of this process, using scans of offices of our department, we were able to reproduce a model which is very close to the ground truth.