Vision systems for a mobile robot based on line detection using the Hough Transform and artificial neural networks.
Damaryam, Gideon Kanji
MetadataShow full item record
This project contributes to the problem of mobile robot self-navigation within a rectilinear framework based on visual data. It proposes a number of vision systems based on detection of straight lines in images captured by a robot using the Hough transform and artificial neural networks as core algorithms. The Hough transform is a robust method for detection of basic features (Boyce et al 1987). However, it is so computationally demanding that it is not commonly used in real time applications and applications which utilise anything but small images (Song and Lyu 2005). (Dempsey and McVey 1992) have suggested that this problem might be resolved if the Hough transform were implemented with artificial neural networks. This project investigates the feasibility of systems using these core algorithms, and systems that are hybrids of them. Prior to application of the core algorithms to a captured image, various stages of pre-processing are carried out including resizing for optimum results, edgedetection, and edge thinning using an adaptation of the thinning method of (Park, 2000) proposed by this work. An analysis of the costs and benefits of thinning as part of pre-processing has also been performed. The Hough transform based system, which has been largely successful, has involved a number of new approaches. These include a peak detection scheme; post-processing schemes which find valid sub-lines of lines found by the peak detection process, and establish which high-level features these sub-lines represent; and an appropriate navigation scheme. Two artificial neural network systems were designed based on lines detection and sub-lines detection respectively. The first was able to detect long lines, but not shorter (even though navigationally important) lines, and so was aborted. The second system has two major stages. Networks of stage 1 developed to detect sub-lines in sub-images derived by breaking down the original images, did so passibly well. A network in stage 2 designed to use the results of stage 1 to guide the robot’s motion did not do so well for most test images. The networks of stage 1, however, have been helpful with development of a hybrid vision system. Suggestions have been made on how this work can be furthered.