Tesis doctoral de Pablo Fernández Alcantarilla
Nowadays, 3d applications have recently become a more and more popular topic in robotics, computer vision or augmented reality. By means of cameras and computer vision techniques, it is possible to obtain accurate 3d models of large-scale environments such as cities. In addition, cameras are low-cost, non-intrusive sensors compared to other sensors such as laser scanners. Furthermore, cameras also offer a rich information about the environment. one application of great interest is the vision-based localization in a prior 3d map. Robots need to perform tasks in the environment autonomously, and for this purpose, is very important to know precisely the location of the robot in the map. In the same way, providing accurate information about the location and spatial orientation of the user in a large-scale environment can be of benefit for those who suffer from visual impairment problems. A safe and autonomous navigation in unknown or known environments, can be a great challenge for those who are blind or are visually impaired. Most of the commercial solutions for visually impaired localization and navigation assistance are based on the satellite global positioning system (gps). However, these solutions are not suitable enough for the visually impaired community in urban-environments. The errors are about of the order of several meters and there are also other problems such gps signal loss or line-of-sight restrictions. In addition, gps does not work if an insufficient number of satellites are directly visible. Therefore, gps cannot be used for indoor environments. Thus, it is important to do further research on new more robust and accurate localization systems. in this thesis we propose several algorithms in order to obtain an accurate real-time vision-based localization from a prior 3d map. For that purpose, it is necessary to compute a 3d map of the environment beforehand. For computing that 3d map, we employ well-known techniques such as simultaneous localization and mapping (slam) or structure from motion (sfm). In this thesis, we implement a visual slam system using a stereo camera as the only sensor that allows to obtain accurate 3d reconstructions of the environment. The proposed slam system is also capable to detect moving objects especially in a close range to the camera up to approximately 5 meters, thanks to a moving objects detection module. This is possible, thanks to a dense scene flow representation of the environment, that allows to obtain the 3d motion of the world points. This moving objects detection module seems to be very effective in highly crowded and dynamic environments, where there are a huge number of dynamic objects such as pedestrians. By means of the moving objects detection module we avoid adding erroneous 3d points into the slam process, yielding much better and consistent 3d reconstruction results. Up to the best of our knowledge, this is the first time that dense scene flow and derived detection of moving objects has been applied in the context of visual slam for challenging crowded and dynamic environments, such as the ones presented in this thesis. in slam and vision-based localization approaches, 3d map points are usually described by means of appearance descriptors. By means of these appearance descriptors, the data association between 3d map elements and perceived 2d image features can be done. In this thesis we have investigated a novel family of appearance descriptors known as gauge-speeded up robust features (g-surf). Those descriptors are based on the use of gauge coordinates. By means of these coordinates every pixel in the image is fixed separately in its own local coordinate frame defined by the local structure itself and consisting of the gradient vector and its perpendicular direction. We have carried out an extensive experimental evaluation on different applications such as image matching, visual object categorization and 3d sfm applications that show the usefulness and improved results of g-surf descriptors against other state-of-the-art descriptors such as the scale invariant feature transform (sift) or surf. in vision-based localization applications, one of the most expensive computational steps is the data association between a large map of 3d points and perceived 2d features in the image. Traditional approaches often rely on purely appearence information for solving the data association step. These algorithms can have a high computational demand and for environments with highly repetitive textures, such as cities, this data association can lead to erroneous results due to the ambiguities introduced by visually similar features. In this thesis we have done an algorithm for predicting the visibility of 3d points by means of a memory based learning approach from a prior 3d reconstruction. Thanks to this learning approach, we can speed-up the data association step by means of the prediction of visible 3d points given a prior camera pose. we have implemented and evaluated visual slam and vision-based localization algorithms for two different applications of great interest: humanoid robots and visually impaired people. Regarding humanoid robots, a monocular vision-based localization algorithm with visibility prediction has been evaluated under different scenarios and different types of sequences such as square trajectories, circular, with moving objects, changes in lighting, etc. A comparison of the localization and mapping error has been done with respect to a precise motion capture system, yielding errors about the order of few cm. Furthermore, we also compared our vision-based localization system with respect to the parallel tracking and mapping (ptam) approach, obtaining much better results with our localization algorithm. With respect to the vision-based localization approach for the visually impaired, we have evaluated the vision-based localization system in indoor and cluttered office-like environments. In addition, we have evaluated the visual slam algorithm with moving objects detection considering test with real visually impaired users in very dynamic environments such as inside the atocha railway station (Madrid, spain) and in the city center of alcalá de henares (Madrid, spain). The obtained results highlight the potential benefits of our approach for the localization of the visually impaired in large and cluttered environments.
Datos académicos de la tesis doctoral «Vision based localization: from humanoid robots to visually impaired people«
- Título de la tesis: Vision based localization: from humanoid robots to visually impaired people
- Autor: Pablo Fernández Alcantarilla
- Universidad: Alcalá
- Fecha de lectura de la tesis: 21/10/2011
Dirección y tribunal
- Director de la tesis
- Luis Miguel Bergasa Pascual
- Tribunal
- Presidente del tribunal: Miguel angel Sotelo vazquez
- José María Martinez montiel (vocal)
- walterio w. Mayol cuevas (vocal)
- Miguel angel Cazorla quevedo (vocal)