Concept and methodology for automated data preprocessing of object recognition algorithm training

Preparing required data for training object recognition algorithms represents a complex and time-consuming process, that must be avoided especially in industrial environments. The work presented in this paper aims to overcome this challenge through on-line machine learning algorithms, as foundation for further developments and validation. The concept and the developed and validated methodology rely on point clouds resulted from the image processing using a depth camera. The geometry and coordinates of the objects are derived from the point clouds, fact that enables the automation of data preprocessing steps (e.g. manually take the pictures, labelling images), optimizing logistics and production activities.