Modeling and Tracking of 3D Objects with Gaussian Mixture

Hybrid use of grid and point cloud data

Abstract

Objects in spatio-temporal data are represented in the form of grid data where scalar values exist on grid points as well as point clouds. Modelling and tracking of such objects, noting events such as splitting and vanishing, will enable summarizing the phenomena represented by the data and discovery of new rules and effective use of data such as forecasting. Hayashi et al. (2019) developed modelling method via mixture of gaussian based on greedy EM extended for the use of grid data, however, this method fails in the specific case tracking moving objects as they grow, causing anomalous split. To suppress these instabilities, we propose a method to generate more natural pseudo point cloud data in which point clouds are distributed even in the gaps between grids, and to use this together with grid scalar data to stably find a solution without increasing the amount of computation. We also compare our method with a different approach, the variational Bayesian method.