Research On Method Of 3D Object Shape Reconstruction By Image Sequences

This paper mainly studies the method of 3D object shape reconstruction from image sequence that is acquired by only one camera. Because the feature set is adopted to achieve the reconstruction, some efforts are made to research the detection of edge and edge line, matching and tracking of feature points, self-calibration of camera parameters, and reconstruction of 3D objects. The main research results achieved in this paper can be summarized as follows:(1) Pixel energy is first defined by square of pixel grey value. Region energy distribution at both sides in different edge is studied. On the other hand, the ratio of the region energy threshold difference and the region area is studied, too. Therefore, based on analysis the relation of edge density distribution and the ratio, an adaptive multiscale edge detection method is proposed. The experiment results show that the novel method, comparing with the Sobel operator, is more effective, more feasible and more noise-resistant in image detection.(2) An adaptive multiscale edge detection method based on wavelet transform is proposed. In this method, the multiresolution technology is used so that the image can be analyzed from coarse to fine. Therefore, the newly proposed edge detection method outperforms the Sobel operator in both resistant of noise and reducing false edge. In addition, the adaptive multiscale edge line extraction method is also investigated. The line equation and its corresponding position information are obtained by the comer mark region and the least square fitting.(3) An adaptive multiscale tracking method is proposed. Based on the maximum similar point principle, the correlation-based method consists of adaptive multiscale of motion object velocity, adaptive threshold region matching, layered-weighting in the matching region, and updating the template adaptively. Experiment shows that the peak of correlation graph obtained with the method is higher and sharper than that obtained by traditional methods. Due to the smaller target matching region and more distinguished matching center obtained in the method, easier matching and higher matching accuracy are achieved. Finally, a region-layered search algorithm is also presented to match the movable target region. This algorithm can improve the tracking speed. (4) Base on spatial plane and spatial orthogonal line, two self-calibration methods are improved to determine the camera parameter, respectively. It differs from the prevalent spatial plane calibration method and spatial vanishing point of orthogonal constraints. Since the focal-length variation between current frame and previous one is considered, the method agrees well with the real application model. In addition, a method of determining second-order radial distortion coefficients based on collinear points is also presented. Comparing with previous methods, this method is much simpler.(5)Based on the camera's intrinsic and external parameter that can be computed by the fundamental matrix and epipolar line, a method for solving the relationship 3D spatial point for 3D object reconstruction is improved. The method is different from previous ones in that it views the focal length as unchanged. In addition, the method for 3D line reconstruction from line equation and shoot plane equation is presented respectively. When there are some lines in image sequence, it is more stable to reconstruct 3D line by 3D line reconstruction than 3D point reconstruction. Compared with line equation, shoot plane equation is used more conveniently to reconstruct 3D line.(6) A method of 3D disfigurement object reconstruction is presented in this paper. The method consists of filling up disfigurement data of object with average of the row and column most optimum curve fitting, solving 3D spatial coordinates and depth coefficient by camera self-calibration considering focal length changed. In addition, a method based on threshold is also improved for comer point detection. Comparing with the existed methods of filling up disfigurement by zero values, our proposed method is more effective in 3D disfigurement object reconstruction.