| Stereo matching is one of the research hot topics in current computer vision, and itis also the most important and difficult step in machine vision. Obtaining large amountof three-dimensional information, difficult process of extracting and representing thefeature points and inappropriate matching results will cause images matchinginaccurately which can directly limit the further development of stereo visiontechnology. So it is significant to study more efficient and fast stereo matchingalgorithm.Based on the large amount of references, the key technologies including cameracalibrating, feature extracting, stereo matching and information capturing are studiedwith the basic framework of Marr vision theory in this paper. Details are as follows:Firstly, Zhang’s calibration method for camera calibration is adopted associating withthe principle of camera calibration with linear model; Secondly, one modified SIFT(scale invariant feature transform) algorithm which extracts Harris corners in themulti-scale space of DOG is developed with the basis of SIFT in stereo matching. Theimproved method extracts Harris corners used as feature points and then makes sure themain orientation for each point. Calculating the64-dimensionl feature vector for eachpoint, matching images by Euclidean distance between the same pointof two images are the following step. Then the parallax of space objects and thecalculating method of the deep information are deduced through the three-dimensionalcoordinates of calculation principle and conversion relationship between image pixelcoordinate and the world coordinate. The last step is to build the hardware platform toachieve the recovering process about object deep information. Experimental results of image matching demonstrate that the new algorithmimproves the significance of the shape of the extracted feature points and gets a bettermatch rate with96%. At the same time the time complexity is reduced by27.8%. Thisalgorithm has a certain practicality. |
PDF Full Text Request