A probabilistic approach for object recognition in a real 3-D office environment

Abstract

The scenario used focuses on object recognition in an office environment scene with the goal of classifying office equipment that is located on a table. The recognition system operates on three-dimensional point-clouds of objects on a loosely covered table where no previous information about the precise position of the table is given. As the point-clouds do not cover the complete objects and the data is noisy, especially for smaller objects a robust detection of special features is difficult. The workflow employed is a three step process: In a first step the table plane is detected and the point clouds of the objects are extracted from the surface. In the second step an object-oriented bounding-box is calculated to get the geometric dimensions, i.e. the properties measured. During a learning phase these simple features are used to calculate the parameters of Bayesian networks. The trained networks are used in the third step, i.e. the classification step. The dimensions of an unknown object form the input for a Bayesian network that yields the most probable object type.