A Study On Robustness Improvement Of Non-rigid Strcuture From Motion

As an important branch of computer vision,the task of three-dimensional(3D)reconstruction is to recover the 3D objects or scenes from two-dimensional(2D)images.3D reconstruction has been widely used in the various applications.As an effective approach of 3D reconstruction,non-rigid structure from motion(NRSFM)can simultaneously recover the 3D shapes of an object and the relative camera motions by utilizing 2D feature points of an image sequence.Nevertheless,a further improvement is still necessary for the accuracy and robustness of the existing NRSFM algorithms,because of the inherent uncertainty of non-rigid motion and the lack of prior information about the 3D shape deformations.In this thesis,the following three work are carried out for the robustness improvement of NRSFM based on some relative constraints.(1)A local deviation constraint based colunn space fitting method is proposed to make the estimation errors of different feature points to be relative uniform.Besides the overall estimation errors,a local deviation constraint is devised by using the variance of reconstruction errors to construct an effective 3D shape estimation model.Furthermore,an augmented Lagrangian multiplier(ALM)iterative algorithm is presented to optimize the constructed estimation model.In additional,the feasible solutions of estimation model and the convergences of the model parameters are analyzed in detail for the model optimization process.The proposed method can achieve a good estimation performance and relatively uniform estimation error for different feature points by utilizing both the overall estimation errors and the local deviation.The experimental results and analysis verify that the estimation model can effectively improve the accuracy of the algorithm and depress the noise.(2)We propose a consistency constraint based Procrustean Markov Process(PMP)model and an accelerated expectation maximization(AEM)algorithm.A consistency constraint based initialization step is designed to suppress the noise for the PMP model.In addition,an accelerated expectation maximization(AEM)algorithm is developed to optimize the PMP estimation model and improve the convergence speedof the model.Experimental results on several widely used sequences demonstrate the effectiveness and feasibility of the algorithm.(3)A local structure based Kermel Shape Trajectory Approach(KSTA)estimation algorithm is proposed for an object with relatively complex and large deformations,such as stretching and bending.First,the object is divided into a large number of local trajectory groups with similar trajectories.Each group is independently used as the input of the KSTA model based weaker estimator.For each weak estimator,a feature vector selection algorithm based rank selection strategy is developed to automatically choose the approximately optimal rank parameter.Finally,a final estimation result for each 3D shape is derived from the output of the KSTA-based weaker estimators by solving a sparse 11-norm minimization model.In general,for the proposed three methods,the robustness of the NRSFM model is significantly improved by reducing the estimation deviation of different feature points,or alleviating the influence of noise and complex deformations.