The term positioning plays a major role not only for specific measurement purposes but also for everyday applications. The problem of positioning in outdoor environment is solved by Global Navigation Satellite Systems (GNSS) to the greatest possible extent. However, localization based on satellite signals fails inside buildings. Therefore, a new field of research has arisen, where alternative methods for positioning within buildings are examined.
In this thesis, a mobile optical sensor system that is based on evaluation methods in the field of computer vision and photogrammetry is designed, which should allow preferably accurate and reliable positioning in a building, using low-cost devices. A cell phone camera is used as sensor system to provide the system for a broad range of users. In this case, the position is determined using a single frame, which contains the projection of an object with four reference points. The object coordinates of the reference points are known. To get the position of the camera according to the object, the object has to be detected and classified in the picture at first. Finally, the image coordinates have to be extracted to determine the position.
First, existing indoor positioning systems (IPS), including optical indoor positioning systems (OIPS), are categorized according to their methods, their accuracy potential, how they depend on the infrastructure and their economy. The image-based system is also classified with respect to these criteria and its capability for indoor positioning is proved in comparison to different IPS.
The accuracy of an OIPS is determined by the configuration of the constellation, the stability of the camera and the quality of the measurements. These factors have not been studied in detail in existing, low-cost OIPS, so that their accuracy potential can not be declared safely. Therefore, in this thesis, the effects of these factors on the position solution are analyzed, using statistical analysis, before designing an own OIPS.
The effect of the configuration can be reduced by appropriate algorithms. Therefore, three different photogrammetric approaches for determining the position of a single image (two 3-point algorithms and a 4-point algorithm) are studied according to their robustness against bad configurations. The approach that is most robust versus configuration, is tested on its behavior to changes in the internal orientation of the camera as well as to measurement noise. So it is estimated a priori, which stability is required for the cell phone camera and how exact the detection of geo-referenced objects in the image has to be.
For system calibration, two methods, single- and multi-view calibration, are compared. Since the parameters of the internal orientation of a camera can be determined more stable in case of the multi-view calibration, this method has been used to determine the parameters of the camera
and to test them according to their stability.
The measurement data, more precisely the points refering the image to the object coordinate system, have to be determined from one single image. Therefore, two new automated methods for classifying the object coordinates of imaged doors (door edges) and for extracting the door points in the image coordinate system are developed and implemented. The quality of the position largely depends on these measurements. Therefore, the two classification methods are compared according to their robustness by using statistical tests. Furthermore, the accuracy of the extraction of the image coordinates is analyzed. In 68% of the cases the doors are classified correctly. With a mean deviation of position in the image points of 0,02 mm, a 3D-accuracy of 1m can be achieved even in bad configurations.
For practical use, the most efficient classification method along with the extraction of the image coordinates and the presented 4-point algorithm (Killian, 1955) is implemented as a mobile application. Further test measurements are performed with the system to evaluate its practicality and to analyze potential areas of application.
In most of the cases, the theoretical estimations can be approved for the application in practice. The results of the test measurements show that the 3D-position can be determined with a mean deviation of position of 0,35 m. | English |