The increasing complexity of today’s mobile networks, due to the coexistence of three generations of mobile communication networks, as well as the traffic growth as a result of the intense use of mobile internet, give origin to daily problems related to optimization and automation of this complex ecosystem. The knowledge of the users’ locomotion method would allow to improve the mobile networks structures planning, the services provided and even the possibility of target marketing.
Therefore, this dissertation presents a framework to develop a Human Activity Recognition (HAR) system to classify the locomotion methods of the users. It does so by analysing the mobile networks traces collected with several smartphones. The labelled data collection followed a well-defined protocol. For the feature extraction was used several temporal overlapping windows over the data, followed by feature selection to find the best subset of features. Lastly, the experimental study of ten supervised learning algorithms was performed to determine the best classifier for the problem.
Evaluation of the final HAR system using 40 hours of mobile network traces collected by six users with different Android smartphones and Random Forest (RF) as the classifier’s algorithm shows that it can achieve an average accuracy of 66% in discerning between eight activities: stationary indoor, stationary outdoor, walk indoor, walk outdoor, run, subway, train and car. This system supports the three generations of mobile networks and can be
extended to the operators terminal data.
Author: Rúben Cunha
Type: MSc thesis
Partner: Faculdade de Ciências e Tecnologia da Universidade Nova de Lisboa