Please use this identifier to cite or link to this item:
|Title:||Vision based obstacle detection for all-terrain robots|
|Authors:||Guedes, Magno Edgar da Silva|
|Publisher:||FCT - UNL|
|Abstract:||This dissertation presents a solution to the problem of obstacle detection in all-terrain environments,with particular interest for mobile robots equipped with a stereo vision sensor. Despite the advantages of vision, over other kind of sensors, such as low cost, light weight and reduced energetic footprint, its usage still presents a series of challenges. These include the difficulty in dealing with the considerable amount of generated data, and the robustness required to manage high levels of noise. Such problems can be diminished by making hard assumptions, like considering that the terrain in front of the robot is planar. Although computation can be considerably saved, such simplifications are not necessarily acceptable in more complex environments, where the terrain may be considerably uneven. This dissertation proposes to extend a well known obstacle detector that relaxes the aforementioned planar terrain assumption, thus rendering it more adequate for unstructured environments. The proposed extensions involve: (1) the introduction of a visual saliency mechanism to focus the detection in regions most likely to contain obstacles; (2) voting filters to diminish sensibility to noise; and (3) the fusion of the detector with a complementary method to create a hybrid solution, and thus, more robust. Experimental results obtained with demanding all-terrain images show that, with the proposed extensions, an increment in terms of robustness and computational efficiency over the original algorithm is observed|
|Description:||Dissertação apresentada na Faculdade de Ciências e Tecnologia da Universidade Nova de Lisboa para obtenção do grau de Mestre em Engenharia Electrotécnica e de Computadores|
|Appears in Collections:||FCT: DEE - Dissertações de Mestrado|
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.