Dissertation > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device

Study on Depth Estimation of Zoom Images with a Single Camera

Author ChenFuGuo
Tutor GaoHongWei
School Shenyang University of Technology
Course Detection Technology and Automation
Keywords Zoom Images Depth Estimation Monocular Stereoscopic Vision SIFT Feature
CLC TP391.41
Type Master's thesis
Year 2012
Downloads 50
Quotes 0
Download Dissertation

Recovering3-d depth from images is a basic problem in computer vision, and hasimportant applications in robotics, scene understanding and3-d reconstruction. To accomplishthe depth estimation of the zoom images, it is necessary to explore the calibration techniqueof the manual zoom camera first. This paper includes the calibration and analysis of cameraparameters, investigation the influence of the characteristic of the manual zoom lens, and astudy of the stability of the parameters of calibration. Therefore, the foundation of the depthestimation of the zoom image is established. It is found that the offline calibration can be usedto get stable focal length and distortion coefficient from analyzing the results of ourexperiments. Although the principal points that were calibrated cannot be used in depthestimation, the zooming center of the zoom image can be utilized. The zooming center by theleast-squares method is obtained. Finally, the optical center displacements of different focusare calibrated. The experiment results show that the differences of the focal lengths are notequal to that of the object distance, and there is a significant difference between those two.Therefore, thick lens model is better in describing the camera zoom lens.On the basis of the Scale Invariant Feature Transform (SIFT) algorithm, the improvedSIFT algorithm is proposed. By using the geometric constraint conditions of zoom image, asmall amount of the ideal matching points are found out. The SIFT features of those idealmatching points are analyzed. Most of the pairs of the ideal matching points are got becauseof these features. Then, by combining the epipolar constraint of the zoom image and the SIFTfeatures of the ideal matching points, the error matching points removal algorithm which isbased on epipolar line distance is proposed. Moreover, the real-time calibration of the zoomcenter is realized. The experiment results show that results of the algorithm is stable. Thisnew algorithm is effective, and good for the removal of the error matching points and gettingthe matching points which have smaller radical angle, and is very helpful in the depthestimate calculation of the zoom image. Finally, the principles of the depth estimation areproposed, and some simple experiments of the depth estimation under the constructed sceneare made to test those principles.

Related Dissertations
More Dissertations