Human localization and activity classification by machine learning on Wi-Fi channel state information

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Sähkötekniikan korkeakoulu | Master's thesis

Date

2020-08-17

Department

Major/Subject

Photonics and Nanotechnology

Mcode

ELEC3037

Degree programme

Master’s Programme in Electronics and Nanotechnology (TS2013)

Language

en

Pages

81+3

Series

Abstract

Devices communicating via Wi-Fi adjust subcarrier correction coefficients in real time. The stream of correction coefficients for all subcarriers is called channel state information (CSI). The latter can be used for human body sensing, in particular current activity and location. The thesis aims to create a robust, environment-agnostic activity classifier. In other words, a neural network (NN) trained to recognize and classify human actions in one location should not dramatically lose prediction capability if transferred to another. This purpose has been achieved in three steps. First, for a neural network to abstract from a particular environment, a diverse data has to be collected. Therefore, a dedicated laboratory equipment to automate a Wi-Fi access point (AP) physical movement and rotation has been developed and constructed. After data for training NNs has been collected, the environment-specific information has been cleaned by classic signal processing algorithms. Finally, a dedicated neural network architecture adjustments have been implemented. Altogether, the goal of environment agnostic classification for the target "sit down", "stand up", "lie down", and "unlie" activities is achieved. However, classification accuracy depends on the similarity between train and test human subjects. The work argues that activity classification and localization tasks have orthogonal goals and focus on different aspects of CSI information. In particular, activity classification NN is interested in ongoing physical movement features and should work independently of the human location. On the contrary, localization NN should ignore subject activities and infer only positioning. Therefore, these two tasks are separated into standalone NN architectures. Since environments such as apartments can be very different from each other, it is assumed that training a universal NN localizer is not possible. Since the localization NN needs to be re-trained for each particular environment, its virtue would be an inexpensive training cycle. Tho achieve this goal, localization NN is substantially reduced in size from 58.8 to 0.4 million parameters.

Description

Supervisor

Särkkä, Simo

Thesis advisor

Honkala, Mikko

Keywords

Wi-Fi, channel state information, convolutional neural network, environment agnostic, activity classification, localization

Other note

Citation