Research On Performance Enhancement Of 3D Reconstruction Of Small Size Image Sequence

Three-dimensional reconstruction is an important branch of computer vision.In some practical applications,such as customs,airports,etc.,the number of people to be dealt with is very large.For a single individual,the images storedin the information database may be relatively small,i.e.small-size image sequence.For a small-size image sequence,the feature points between two adjacent frames may vary greatly,i.e.these two frames are non-smoothing.The accuracy of 3D reconstruction generally decreases significantly,when the frame number is small,or two adjacent frames are non-smoothing.To address the above issues,two works have been carried out in thisthesis.(1)For a single 2D image,a3 D shape estimation approach is proposed via the constraints to adjust the sparsity of the coefficients.Given the training data,a sparse representation model with elastic network is first constructed to extract shape bases.In the sparse representation model,the elastic network,i.e.L1-norm and L2-norm,is adopted to adjust the sparsity and scale of the coefficients.Secondly,a penalty least squares model is established for 3D shape and motion estimation.Two constraints are considered in the model,i.e.the orthogonal constraint of the transformation matrix,and the similarity constraint between the two-dimensional least squares and the shape bases.Finally,the estimation model is optimized by the augmented Lagrange multiplier iterative algorithm to derive 3D shapes.The experimental results on the well-known CMU image sequence demonstrate the effectiveness and feasibility of the proposed algorithm.(2)A virtual sampleconstruction method is proposed based on the Gaussian mixture model(GMM),to improve the smoothness of image sequence and the 3D reconstruction accuracy.First,for two adjacent frames,a GMM of point set registration is established for the point sets according to the corresponding relationship.During the optimization process of GMM,the transformed point sets at different iteration steps are successively selected as the new samples between two frames,to enhance the smoothness of two adjacent frames.Then,a reconstruction model based on probability distribution is established by using the given image sequence and the new samples.Finally,3D shapes are estimated via theexpectation-maximization algorithm by optimizing the model parameters.The experimental results demonstrate that the proposed method has the higher estimation accuracy than several existing algorithms.