Predicting observation results from a new viewpoint using deep learning for mobile robot viewpoint planning

Abstract

In this paper, we describe a viewpoint planning method by predicting observation results using deep neural networks. We use two networks for viewpoint planning. The first network predicts the size of a newly observed area, while the second one predicts an updated map after observation. We develop a look-ahead viewpoint planning using these results. We compare the proposed method with a geometrical method and a method which uses the first neural network to consider only the current state. The proposed method is the most balanced method among them.