Dimensionality Reduction Through Multiple Convolutional Channels for RSS-Based Indoor Localization

Abstract

Dimensionality reduction is an important task for Wi-Fi-based indoor localization (IL). Most such techniques do not take into account realistic data collection issues such as the presence of outliers or inconsistent fingerprint instances. These fingerprints either represent a class boundary or an outlier. Instance hardness is a measure that better characterizes such instances. Accordingly, in this work, our contribution is to propose a convolutional autoencoder-based dimensionality reduction approach that works on the basis of feature transformation and instance hardness. The encoding process of the data input involves a two-channel representation of a fingerprint dataset that holds the normalized RSS and an instance hardness measure, that is, a k-disagreeing score. The inclusion of the k-disagreeing score into the training pipeline is made with the objective of injecting instance importance for training using 1-D CNN architectures for classification. The experimentations were performed on three benchmark datasets and a collected dataset. The proposed pipeline is found to yield an accuracy of more than 97% with error deviation ranging from 2.2– $2.37{m}$ which is quite acceptable for any localization system.