A Comparative Analysis of Models Based on Convolutional Neural Network for WiFi Fingerprinting

Abstract

Location-based services are nowadays more popular due to their wide applications in human daily life. One of the most famous services is location tracking which aims to determine the position of a person or an object in geographic coordinates. For indoor positioning, WiFi Fingerprinting-based systems are receiving much attention due to the utilization of the building infrastructure. However, the instability of the WiFi signals makes it difficult to ensure the positioning performance. Thus, to improve positioning accuracy, deep learning techniques are being used. In this paper, different models that are built based on Convolutional Neural Networks (CNNs) and stacked autoencoder (SAE) are implemented and analyzed in various public datasets with different structures. The SAE is used to extract major features from the WiFi data and to reduce the training time of the CNN-based models. The experimental results reveal that the combination of AE with only two layers of CNN obtained the best results in most of the compared datasets with an average error of only 3.85 m, which is smaller than other models of up to 16.49%. In addition, the prediction time of the above model is very competitive to others.