Introduction

This project aims to detect the occupant in a certain space, in which the space is defined based on its social nature. For example, it is meaningful to detect the occupant in a kitchen, rather than in a random location along the corridor. Such meaningful information allows us to trigger corresponding IoT application related to the social nature of that space.

Dataset

The dataset were collected in Rio Hotel, Macau, from June 12-15 2017. The dataset is available on our Github repository. Two different devices (i.e., Asus Zenphone Deluxe and Samsung Tablet) were used to collect the data, detailed description of the data is available in this link.

The figure below indicate the all the training samples in 3D plot.

Data Processing

We had applied (1) denoisining autoencoder, (2) contractive autoencoder and (3) stacked autoencoder to process the data. The motivation to apply the denoising autoencoder is that the collected data is inherently noisy due to the uncontrollable environmental factors, such as multipath and shadowing. We also apply the contractive autoencoder to deal with the always changing RSS values. Since the dimension of the fingerprint vector is small, we apply a fixed dictionary to achieve fingerprint expansion. The resultant fingerprints at each zone is shown below. We used 2D plot for easy visualization. More details are available in our paper.