Topology Preserving Input Image for Convolutional Neural Network Based Indoor Localization

International Conference on Indoor Positioning and Indoor Navigation, pp. 1–8

Abstract

Convolutional neural networks (CNNs) have made a significant impact on problems such as image classification. By learning feature maps via convolutional layers, local features of the images are extracted. Recently, CNNs have also been applied for fingerprinting based indoor localization that use the received signal strength (RSS) of WLAN access points (APs) to construct the fingerprints. The input images are created from the RSS vector by arbitrarily assigning APs to pixel positions of the 2D-image. Local information is lost by this procedure, which contradicts the prior learned by CNN that close pixels are more related than distant ones. In this work we try to preserve the topology of the APs within the constructed 2D-input image. We propose an unsupervised method to estimate pairwise relative AP distances from unlabeled fingerprints. We develop an algorithm to generate an AP-image mapping that assures highly related APs will appear close in the resulting image. On the public fingerprinting benchmark dataset from UJI, we empirically show that the topology preserving AP-images (TPAI) significantly benefit classification accuracy of various recently proposed CNN architectures.

Authors 2

1. Marius Laska Aachen Geodetic Institute Chair for Computing in Civil Engineering & Geo Information Systems

   RWTH Aachen University

   Affiliation as printed

   Geodetic Institute and Chair for Computing in Civil Engineering & Geo Information Systems, RWTH Aachen University, Aachen, Germany

2. Jörg Blankenbach Aachen Geodetic Institute Chair for Computing in Civil Engineering & Geo Information Systems

   RWTH Aachen University

   Affiliation as printed

   Geodetic Institute and Chair for Computing in Civil Engineering & Geo Information Systems, RWTH Aachen University, Aachen, Germany