InPosNet: Context Aware DNN for Visual SLAM

Abstract

This paper introduces a novel approach to accurately localize a subject in indoor environments by using the scene images captured from the subject’s mobile phone camera. The objective of this work is to present a novel deep neural network (DNN), called InPosNet, that generates a concise representation of an indoor scene while being able to distinguish between their inherent symmetry. It also enables the user in real time distinction between the images of the same location but captured from different orientations, thereby enabling the user to detect the orientation along with position. A localization accuracy of less than 1 meter from ground truth is achieved and enumerated through the experimental results. The novel DNN presented in the work is motivated by MobileNetv3-Small [2], followed by PCA based feature space transformation. PCA helps in feature space dimensionality reduction and projection of query images onto an optimally dense subspace of the original latent feature space. The goal is to present a vision based system that will have the ability to be used for indoor positioning, without any need for additional infrastructure or external hardware.