Realistic images generation for training artificial neural networks in robot navigation problem

Abstract

The problem of obtaining training dataset for setting weights of neural network designed for indoor detection of doors is considered and solved on the particular example. A method for generating realistic synthetic data is developed. The method involves replacing a priori known target objects on digital images with new reference objects that were obtained by projective transformation of reference objects. The method is designed to obtain training dataset for training and testing of artificial neural networks, which will be used in the mobile robot control system to solve autonomous navigation problem. The effectiveness of the proposed method was confirmed experimentally.