| Computer Vision has been evolving as a multidisciplinary subject. Its contours blend into those of artificial intelligence, robotics, signal processing, pattern recognition, control theory, psychology, neuroscience, and other fields. Its research area and application area spread widely. Recently, price of the video devices is going down, and performance of computer is going up. All of these promote the research and application of computer vision to a new level. Studying on computer vision has important significance not only on science but also on practical application.Computer vision is a synthesis and cognition science of 3D world from one or more digital images. Generally, the key techniques of a computer vision system is composed of five parts, which are video capture, camera calibration, image pre-processing and feature extraction, stereo correspondence and 3D reconstruction. This dissertation deals with these key problems in computer vision, and presented some new ideas, approaches and come to some valuable results.The original objects processed in computer vision are digital images captured by digital video devices. System structure and technical details in video capture are studied. Based on Windows platform, a real time video capture system is developed by accessing to buffer directly. It can satisfy the need of real time vision system well.As one of the basic tasks in computer vision, camera calibration determines the internal and external parameters of a vision system. In this dissertation, based on modified pine-hole camera model with one order radial lens deformation, a new linear multi-step method for camera calibration is developed. With some appropriate transformation and arrangement, the camera parameters can be calculated by solving these linear equations sequentially only. The running result of this new method shows that the algorithm is simple and efficient, and the precision is good enough. A image distortion correction method for radial lens distortion with segment slope is also described. Both simulated image and real image running results show that this correction method is robust and accurate.There has been growing interest of wavelet-based denoising schemes and edge detection techniques recently. Such popularity is mainly due to that wavelet provides an appropriate basis for separating noise signal from image signal and makes noise and edge show different attributes. A denoising algorithm based on wavelet transform with redundant representation is proposed, which can not onlymake improved denoising performance, but also suppress the Gibbs phenomena. A method of edge detection based on wavelets transform is also presented, in which a self-adapted method for selecting the threshold of edge chain evenness is adopted to detect edge well.Stereo correspondence is to determine that an item in one image corresponds to another item in other images of the same scene. Based on edge points, a two-stage stereo matching method with two thresholds is proposed.At last, a real time vision recognition system of MIROSOT is developed to recognize all vehicle robots and football. In this system, the rapid processing speed and high recognizing accuracy are needed. To meet such requirements, a recognition algorithm based on object projection is presented. The vision recognition is processed with two stages. In the first stage, a grid search scanning to find an object point inside an unrecognizing object is processed; then a following precision analyzing and recognizing procedure based on object region projection is processed for determining the robots and football's position and also the robots' orientation in the second stage. The recognition algorithm is robust in treating noisy image captured by CCD, even in the situation while multiple vehicle robots are closely adjacent each other. The algorithm is also very efficient, it takes only 2ms running in PII350 PC for recognizing and identifying all objects in a captured frame.
|
PDF Full Text Request