Robust Cuboid Modeling from Noisy and Incomplete 3D Point Clouds Using Gaussian Mixture Model

A cuboid is a geometric primitive characterized by six planes with spatial constraints, such as orthogonality and parallelism. These characteristics uniquely define cuboid. Therefore, previous modeling schemes have used these hard which narrowed the solution space for estimating parameters of However, under high noise occlusion conditions, may contain only false or no solutions, called an over-constraint. In this paper, we propose robust method point clouds conditions. The proposed estimates using soft which, unlike do not limit space. For purpose, represented Gaussian mixture model (GMM). distribution each surface owing to assumed be model. Because face cuboid, GMM shares satisfies regardless occlusion. To avoid over-constraint in optimization, constraints are employed, expectation GMM. Subsequently, maximized analytic partial derivatives. was evaluated both synthetic real data. data were hierarchically designed test performance various data, more dynamic than follow assumption. acquired light detection ranging-based simultaneous localization mapping actual boxes arbitrarily located indoor experimental results indicated that outperforms terms robustness.