Research Of Corner Detection And Matching Based On Multi Scales

Intelligent robots have been used more widely with the development of human society. In order to let intelligent robots serve people better, making robots have the perceptive ability of eye and the decision-making ability of brain is the aim that many researchers made great efforts to achieve. The purpose of vision researching is that letting robots have the perceptive ability similar to human vision.3D reconstruction is one of the main contents that robot vision includes, and corner detection in2D image and matching based on Binocular stereo vision are the cores of3D reconstruction. So the paper makes special researches on the problems of corner detection and matching.Firstly, the paper introduces three kinds of classical algorithms of corner detection, and completes the experiments of corner detection separately based on these three algorithms. Besides those above, the paper also gives brief explanations of traditional algorithms of matching.Secondly, classical algorithms of corner detection are improved due to these drawbacks, and a new algorithm of corner detection which is named "multi scales" algorithm of corner detection’ is proposed. In the algorithm the image is classified by its complexity at first. If the environment of the image is simple, the algorithm will detect the edges of object, and the edges are smoothed by Gaussian function at different scales, algorithm judges corners’ existing through edges’ difference after being smoothed. If the environment of the image is complex, the image will be smoothed by Gaussian function and detected by Harris algorithm again and again, and real corners are distinguished from noise according to the feature that corner’s stability is better than noise. Through some experiments, it is proven that multi scales" algorithm of corner detection is better than other classical algorithms of corner detection.Thirdly, a new algorithm of matching based on scale, distance and rotation is proposed because traditional algorithm of matching based on similar measure directly has many drawbacks. The algorithm makes the first match through using similar frequency domains, and matching candidates need to be estimated based on scale, distance and rotation, then right matches are distinguished from wrong matches by function values. Through some experiments, it is proven that the algorithm of matching based on scale, distance and rotation is better than traditional algorithms in matching corners.Finally, achievement of thesis is summarized, ideas about corner detection and matching are introduced, and vein research forecast is also proposed.