| Résumé : |
(Auteur) A fundamental task of an autonomous mobile robot is the ability of self-localization in its environment respectively in a map of it, available to the robot. Many applications require a localization that is as precise as possible. Thereby, a decisive factor is the accuracy but also the completeness of the map. The generated map can usually be seen as a necessary side-product. When considered from a surveying point of view, it is brought more into focus. The question is to what extent the spatial robot data can fulfill certain mapping requirements in terms of accuracy and completeness in a detailed enough manner to be useful to human users. Precise models of indoor environments are very useful in both public and private sectors. But since their procurement involves a great deal of effort, an automatical generation of indoor models is desirable.
The following thesis should make a contribution to this and tries to compose the techniques that are necessary to deliver interior models at the push of a button. Therefore, a mobile measuring system was designed, that is able to carry out complete and precise measurements of indoor environments. The system was build up in two stages. A mobile platform is equipped with a low-cost laser scanner in the basic stage. This build-up allows a precise exploration of indoor scenes in 2D. In an expansion stage the installation of a digital camera leads to an additional 3D reconstruction based on photogrammetric techniques.
The accuracy and quality of robotic mapping is primarily dependent on the sensors that are used. The software process to register the collected data in a common coordinate frame and to create a holistic map of the environment has an additional influence. In the field of scanning systems scan matching techniques or probabilistic filter approaches are used. In the case of 3D surveying, a photogrammetric reconstruction can be done by an estimation of feature points, which are extracted from a photo block, within a bundle adjustment process. Results can be optimized, if requirements and restrictions of these techniques are regarded in the data acquisition process, when the robot proceeds the exploration. A crucial aspect is the spatial sensor placement at a certain point of time in this process. In the field of robotics, sensor placement is controlled by positioning strategies, which normally are of overriding importance to all other processes. The majority of known positioning strategies have the primary aim to optimize the efficiency of the exploration, so that every measurement provides as much new spatial information as possible. Requirements of the data registration technique are neglected in contrast. This thesis presents positioning strategies for scanning 2D measurements as well as photogrammetric 3D measurements, which try to maximize the accuracy of the collected spatial data. A scanning 2D measuring system, which is able to explore previously unknown indoor environments and generate precise floor plans is presented in the first part of this thesis. The system iteratively visits measurement poses defined by a positioning strategy. 2D point clouds, collected at various positions, are transformed into a common coordinate system by the use of a scan matching technique. The latter takes advantage of the characteristic manifestation of office environments and extracts planar segments from the measured point clouds. Minimizing the sum of perpendicular distances to these segments, points of a new scan are transformed into the existing coordinate system with high accuracy. Precision and robustness are improved through iterative parameter refinement. The measuring system uses a positioning strategy, which is based on the global assumption that the environment can be described as a collection of line segments. Since segment ends indicate data gaps, exploration is pushed until their observation is complete. All accessible measuring positions, represented in an occupancy grid, are evaluated in terms of their explorative benefit by the strategy using a cost function. Exploration is stopped as soon as every section is observed with a desired resolution.
The second part of this work presents a positioning strategy to enable the recording of photo blocks that are suitable for a photogrammetric reconstruction. In the run-up of data acquisition possible pose configurations are determined using accuracy estimation. The assumption of plain environments allows a limitation to a 2D search problem regarding the choice of possible camera poses, whereby the combinatorial variety is reduced. Initial information of pose estimation provides a 2D map, generated by the system structure that was presented in the first part of this work. For predefined sections of the environment, pseudo-random pose constellations are derived iteratively from the map and compared with each other using a cost function. The cost function helps to predict the variances resulting from a bundle adjustment. Therefore the functional model of the bundle adjustment has to be projected on the R2 in a way that a geodetic network consisting of direction measurements remains. The size and the shape of the resulting error ellipses allow conclusions and a comparative consideration regarding the quality of camera pose candidates.
An essential part of the work is the empirical analysis of the systems, to evaluate their performance and the quality of the resulting spatial data. Various experiments in real indoor environments show that developed measurement methods can be applied in practice. In different sets of experiments initial conditions are varied to find out their influence on the measurement process or the result. In order to achieve reliable results, reference models of the experimental environments were created by the use of a total station.
In the case of scanner measurements, experiments show that the developed system is able to explore and measure also complex interiors. An examination of the point clouds show that the achieved accuracy comes up with surveying demands. On this issue, the presented technique outplays conventional measuring equipment. However, additional modeling shows that mainly fine structures of the environment are displayed wrongly or are even lost completely. Also the 3D measuring strategy is demonstrably superior to existing techniques. The purely passive technique leads to sparse point clouds, not dense enough to derive detailed environment models with the corresponding software. |
Ableitung von Bewegungsstrategien zur automatisierten, vollständigen Vermessung von Innerraumszenen auf autonom navigierender Plattform (2012)
