| This dissertation describes our investigation on some algorithm and realization technique of computer vision. Research fruits mainly include image feature detection and realization, the real-time detection algorithm of active contour, segmentation algorithm of moving objects from sequential images and architecture design for motion objects detection etc. The motivation of this dissertation is how to find simple and effective algorithms and suitable related architectures, so as to combine the theory, algorithm and architecture into application. The contributions of this dissertation is listed below.Aiming at existing defects of traditional beeline detection by Hough Transformation (HT) arithmetic, such as large storage space consuming, imprecise beeline detection effect and information loss of image feature pixels etc, we suggest two new beeline detection methods. (1) Based on infinite symmetry exponential filter and its recursion arithmetic, new beeline detection methods can get exact direction of image edge pixels in advance, so it reduces highly the calculation quantity of the conventional HT, and improves the speed and the precision of beeline detection greatly. (2) The relation of each image feature point and the most evident curve membership is discovered, such relation can be used to reduce redundancy of parameter space, depresses computational effort and suppresses the false accumulator cells etc. The most evident curve membership is then fed to the original image space, the information of image pixels can be recovered, the shorter line can be found, the bad line can also be repaired and the precision of beeline detection can be improved.On the basis of studying MIC corner detector, a new adaptive window interpolation corner detection algorithm is proposed, which has two advantages over MIC algorithm. Firstly, concept of fuzziness is used in this new adaptive window interpolation corner detection scheme. It is superior to the widely adopted algorithm with fixed window interpolation which often causes right corner missing > false corner alarming or imprecise corner location. Secondly, using the character that the edge pixels in oblique line have continuous intensity along the direction, some kind of false corners caused by quantification error are removed easily. Experiment results shows that the probability and accuracy of corner detection are improved obviously.Based on the gradual intensity change of the image, a new redefinition of the regionof USAN and a more efficient and simple method to compute the region area are suggested. And thus a new corner detection algorithm RSUSAN (Redefined SUSAN) is proposed. The experiments show, comparing with classical SUSAN and MIC corner detectors, this algorithm is of simpler process, rapid speed and higher veracity. In addition, in order to deal with the blurred and noisy image, the way of combining adaptive smoothing method with above RSUSAN algorithm is put forward, and the results show such combination has better effect than SUSAN and MIC in corner detection to blurred and noisy image.Basing on studying the principle of active contour model and some solutions to it, a new partial optimal dynamic programming contour detection algorithm is presented and the results comparable to the dynamic programming algorithm are given. Experiment shows such new algorithm is much faster.The segmentation of video sequence into independently moving objects accurately play a very important role in the motion detection of vision and the new video coding standard MPEG-4.0n the basis of studying some recent motion objects segmentation algorithms, a new automatic video sequence segmentation algorithm based on active contour that extracts moving objects is presented. Experiment testifies that the proposed algorithm is of few parameters, robust to noise, best in result of segmentation and with quick speed.At last, in order to testify our algorithm and architecture proposed, we design a dynamic image collection and disposing platform by USB2.0 bus, and experiment...
|
PDF Full Text Request