Abstract:
With the advent of 5G and 6G technologies, WiFi based outdoor positioning and lo calization schemes have found applications in the domains of autonomous driving, IoT
(Internet of Things), vehicular tracking, smart cities etc. Most of these positioning
algorithms are based on supervised learning approaches which require copious amount
of labelled data which is costly. However, the use of Massive MIMO Transmitters and
Receivers have led to the availability of vast amount of unlabelled data. In this re search we propose a baseline DCNN (Deep Convolutional Neural Network) supervised
model that takes CSI (Channel State Information) values, gathered by 8 x 8 massive
MIMO (Multiple Input Multiple Output) antenna array as input for estimating user
positional coordinates. To leverage the information stored in the unlabelled data, we
propose a self-supervised learning approach that consists of a pretext and downstream
task. Pretext task consists of an Autoencoder which extracts the useful features from
the unlabelled CSI of 56 antennas and 924 subcarriers. The trained Encoder from the
pretext task is later utilized in the downstream task to extract features of labelled CSI
samples. The extracted features are utilized as an input to the downstream DCNN for
predicting the position of the user. Both supervised and self-supervised methods are
compared. The proposed self-supervised method yields better accuracy as compared
to the supervised model. Hence establishing that leveraging the unlabeled data using
self-supervised techniques can lead to better performance and accuracy in the domain
of localization.
