This thesis addresses the issues of visual landmark recognition in autonomous robot navigation along known routes, by intuitively exploiting the functions of the human visual system and its navigational ability. A feedforward-feedbackward architecture has been developed for recognising visual landmarks in real-time. It integrates the theoretical concepts from the pre-attentive and attentive stages in the human visual system, the selective attention adaptive resonance theory neural network and its derivatives, and computational approaches toward object recognition in computer vision. / The main contributions of this thesis lie within the emulations of the pre-attentive and attentive stages in the context of object recognition, embedding various concepts from neural networks into a computational template-matching approach in the computer vision. The real-time landmark recognition capability is achieved by mimicking the pre-attentive stage, where it models a selective attention mechanism for computational resource allocation, focusing only on the regions of interest. This results in a parsimonious searching method, addressing the computational restrictive nature of current computer processing power. Subsequently, the recognition of visual landmarks in both clean and cluttered backgrounds (invariant to different viewpoints) is implemented in the attentive stage. This is achieved by developing a memory feedback modulation (MFM) mechanism that enables knowledge from the memory to interact and enhance the efficiency of earlier stages in the system, and the use of viewer-centre object representation which is mimicked from the human visual system. Furthermore, the architecture has been extended to incorporate both top-down and bottom-up facilitatory and inhibition pathways between the memory and the earlier stages to enable the architecture to recognise a 2D landmark, which is partially occluded by adjacent features in the neighbourhood. / The feasibility of the architecture in recognising objects in cluttered backgrounds is demonstrated via computer simulations using real-images, consisting of a larger number of real cluttered indoor and outdoor scenes. The system's applicability in mobile robot navigation is revealed through real-time navigation trials of known routes, using a real robotic vehicle which is designed and constructed from the component level. The system has been evaluated by providing the robot with a topological map of the routes prior to navigation, such that object recognition serves as landmark detection with reference to the given map, where autonomous guidance is based on the recognition of familiar objects to compute the robot's absolute position along the pathways. / Thesis (PhD)--University of South Australia, 2006.
| Identifer | oai:union.ndltd.org:ADTP/267287 |
| Creators | Do, Quoc Vong. |
| Source Sets | Australiasian Digital Theses Program |
| Language | English |
| Detected Language | English |
| Rights | copyright under review |
Page generated in 0.0019 seconds