Um descritor robusto e eficiente de pontos de interesse: desenvolvimento e aplicações

Abstract

At the core of a myriad of tasks such as object recognition, tridimensional reconstruction and alignment resides the critical problem of correspondence. Due to the ambiguity in our world and the presence of noise in the data aquisition process, performing high quality correspondence is one of the most challenging tasks in robotics and computer vision. Hence, devising descriptors, which identify the entities to be matched and that are able to correctly and reliably establish pairs of corresponding points is of central importance. In this thesis, we introduce three novel descriptors that efficiently combine appearance and geometrical shape information from RGB-D images, and are largely invariant to rotation, illumination changes and scale transformations. For applications that demand speed performance in lieu of a sophisticated and more precise matching process, scale and rotation invariance may be easily disabled. Results of several experiments described here demonstrate that as far as precision and robustness are concerned, our descriptors compare favorably to three standard descriptors in the literature, namely: SIFT, SURF (for textured images) and Spin-Images (for geometrical shape information). In addition, they outperfom the state-of-theart CSHOT, which, as well as our descriptors, combines texture and geometry. We use these new descriptors to detect and recognize objects under different illumination conditions to provide semantic information in a mapping task. Furthermore, we apply our descriptors for registering multiple indoor textured depth maps, and demonstrate that they are robust and provide reliable results even for sparsely textured and poorly illuminated scenes. In these two applications we compare the performance of our descriptors against the standard ones in the literature and the state-of-the-art. Experimental results show that our descriptors are superior to the others in