Browsing by Author "Veiga, Ricardo Jorge Martins da"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Pose estimation system based on monocular camerasPublication . Veiga, Ricardo Jorge Martins da; Rodrigues, J. M .F.; Cardoso, Pedro J. S.Our world is full of wonders. It is filled with mysteries and challenges, which through the ages inspired and called for the human civilization to grow itself, either philosophically or sociologically. In time, humans reached their own physical limitations; nevertheless, we created technology to help us overcome it. Like the ancient uncovered land, we are pulled into the discovery and innovation of our time. All of this is possible due to a very human characteristic - our imagination. The world that surrounds us is mostly already discovered, but with the power of computer vision (CV) and augmented reality (AR), we are able to live in multiple hidden universes alongside our own. With the increasing performance and capabilities of the current mobile devices, AR is what we dream it can be. There are still many obstacles, but this future is already our reality, and with the evolving technologies closing the gap between the real and the virtual world, soon it will be possible for us to surround ourselves into other dimensions, or fuse them with our own. This thesis focuses on the development of a system to predict the camera’s pose estimation in the real-world regarding to the virtual world axis. The work was developed as a sub-module integrated on the M5SAR project: Mobile Five Senses Augmented Reality System for Museums, aiming to a more immerse experience with the total or partial replacement of the environments’ surroundings. It is based mainly on man-made buildings indoors and their typical rectangular cuboid shape. With the possibility of knowing the user’s camera direction, we can then superimpose dynamic AR content, inviting the user to explore the hidden worlds. The M5SAR project introduced a new way to explore the existent historical museums by exploring the human’s five senses: hearing, smell, taste, touch, vision. With this innovative technology, the user is able to enhance their visitation and immerse themselves into a virtual world blended with our reality. A mobile device application was built containing an innovating framework: MIRAR - Mobile Image Recognition based Augmented Reality - containing object recognition, navigation, and additional AR information projection in order to enrich the users’ visit, providing an intuitive and compelling information regarding the available artworks, exploring the hearing and vision senses. A device specially designed was built to explore the additional three senses: smell, taste and touch which, when attached to a mobile device, either smartphone or tablet, would pair with it and automatically react in with the offered narrative related to the artwork, immersing the user with a sensorial experience. As mentioned above, the work presented on this thesis is relative to a sub-module of the MIRAR regarding environment detection and the superimposition of AR content. With the main goal being the full replacement of the walls’ contents, and with the possibility of keeping the artwork visible or not, it presented an additional challenge with the limitation of using only monocular cameras. Without the depth information, any 2D image of an environment, to a computer doesn’t represent the tridimensional layout of the real-world dimensions. Nevertheless, man-based building tends to follow a rectangular approach to divisions’ constructions, which allows for a prediction to where the vanishing point on any environment image may point, allowing the reconstruction of an environment’s layout from a 2D image. Furthermore, combining this information with an initial localization through an improved image recognition to retrieve the camera’s spatial position regarding to the real-world coordinates and the virtual-world, alas, pose estimation, allowed for the possibility of superimposing specific localized AR content over the user’s mobile device frame, in order to immerse, i.e., a museum’s visitor into another era correlated to the present artworks’ historical period. Through the work developed for this thesis, it was also presented a better planar surface in space rectification and retrieval, a hybrid and scalable multiple images matching system, a more stabilized outlier filtration applied to the camera’s axis, and a continuous tracking system that works with uncalibrated cameras and is able to achieve particularly obtuse angles and still maintain the surface superimposition. Furthermore, a novelty method using deep learning models for semantic segmentation was introduced for indoor layout estimation based on monocular images. Contrary to the previous developed methods, there is no need to perform geometric calculations to achieve a near state of the art performance with a fraction of the parameters required by similar methods. Contrary to the previous work presented on this thesis, this method performs well even in unseen and cluttered rooms if they follow the Manhattan assumption. An additional lightweight application to retrieve the camera pose estimation is presented using the proposed method.