Object shape estimation from point clouds using mesh-based and volumetric-based shape priors

Abstract

This thesis presents and evaluates two novel methods for the object shape estimation from point clouds. Both methods use prior knowledge of the expected object shape. In the first part of this thesis a novel robust mesh-based shape estimation method is presented. The estimation method uses the prior knowledge from a statistical shape model. The statistical shape model is used to define a similarity error. This similarity error is used to stabilize the estimation procedure. Different functions for a similarity error are evaluated and it is shown that a quadratic similarity error function achieved the lowest error in dependency of the noise in the point cloud. It is shown that the implemented method achieves more accurate results for noisy point clouds than other reconstruction methods like the Poisson, SSD and Wavelet reconstruction. It is shown that the implemented method achieves a smaller reconstruction error for incomplete point clouds than other reconstruction methods like the Poisson, SSD and Wavelet reconstruction. It is evaluated how the weight of the similarity error influences the achieved accuracy. It is shown how a point cloud uncertainty value can be computed to set the weight of the similarity error. In addition the runtime behaviour, the influence of the number of used shapes, the influence of the object pose error and the estimation for missing areas in point clouds is evaluated. In the second part of this thesis a novel volumetric-based shape estimation method is presented. The estimation method uses the prior knowledge from a Gaussian process latent variable model. The latent space of the Gaussian process latent variable model is used for the optimization. Two different similarity errors are introduced to stabilize the estimation procedure. It is shown how the similarity errors can reduce possible distortions on the estimated object surface. It is evaluated how the number of DCT coefficients and the size of the latent space influence the estimated shape. It is demonstrated how different artifacts can occur on the object surface during the interpolation between multiple shapes. It is also shown how these artifacts can be reduced. In addition different failure cases during the interpolation between multiple shapes are evaluated. It is shown how the estimated shape can be used to increase the accuracy and completeness of the measured point cloud. In addition the runtime behaviour is evaluated.