Pregled bibliografske jedinice broj: 56668
COMPUTER VISION, 2000. (elaborat).
Ostale vrste radova, elaborat
Computer vision; pattern recognition; robot vision
The purpose of this report is to provide some ideas and basic principles which can be used to design the software for a vision system. It discusses concepts that are generally applicable in vision systems but it is primarily concerned with the application in robotics. Some very important concepts of computer vision such as stereo-vision, texture recognition etc. are not covered in this text. A control system in general is an agent that performs some task by influencing its environment using it's actuators, according to the observations of the effects of its actions and effects of other factors that influence the environment provided by the sensors. In this text, systems that use signal obtained by visual perception for observing its environment are discussed. There are two main tasks of vision perception in robotics: recognition of objects in the agent's environment and object pose determination, i.e. determination of the positions and orientations of the objects relative to the referent coordinate system of the agent. A structure of computer vision system is discussed. The components of that structure are: image filters, edge detector, edge vectorization algorithm, image feature extractor, part detector and object detector. Image filters their purpose is discussed. As an example Gaussian filter is given. Edge detection is also discussed: a very popular edge detection method proposed by Canny is described. As a extension to edge detection, edge vectorization is discussed. The Gaussian filter, Canny edge detector and an edge vectorization algorithm which uses ideas presented in this text are implemented as components of Robot Vision Library (RVL). RVL is a software platform implemented at Department of Control and Computer Engineering in Automation, Faculty of Electrical Engineering and Computing, University of Zagreb using Microsoft Visual C++. It's purpose is to provide a shell for prototyping and testing of algorithms for image analysis and computer vision systems. A very common 2D shape recognition method - the Hough transform is described. The basic principle of the Hough transform is used in the indexing method for 3D object recognition. The method has two phases: model acquisition during which the descriptions of 3D models are stored in the agent's memory and recognition during which instances of models are detected in images using model descriptions stored during model acquisition. Besides recognizing an object the indexing method also provides the estimate of the pose of the recognized object which can be used for exact pose determination. It is also suggested how the principle of indexing can be used to construct a system that could be able to extract 3D objects from a sequence of images without being previously provided with the descriptions of 3D models.