A Hybrid Framework for Visual Positioning: Combining Convolutional Neural Networks with Ontologies

Abstract

Visual positioning is a new generation positioning technique which has been developed rapidly during recent years for many applications such as robotics, self-driving vehicles and positioning for visually impaired people due to advent of powerful image processing methods, especially Convolutional Neural Networks. Nowadays, deep Convolutional Neural Networks are capable of classifying images with high accuracy rates; however, comparing visual perception by a human being, pure Neural Networks lack background knowledge which is essential for estimating the position through a reasoning process. In this paper we present a hybrid framework for employing ontologies over Convolutional Neural Networks to integrate a knowledge-based reasoning with Neural Networks for taking advantages of capabilities similar to human brain’s functions. The proposed framework is generic so it can be applied to a wide variety of scenarios in smart cities where visual positioning represents added value.