Predicting QoE of live videos based on measurements in the LTE system
Authors
Vourvachis, Markos ; Jäger, Andreas Vembe
Term
4. term
Education
Publication year
2018
Submitted on
2018-06-07
Pages
64
Abstract
Mobilnetværksoperatører har brug for mål, der afspejler, hvad brugerne faktisk oplever. I stedet for kun at bruge netværkscentriske Quality of Service (QoS) værdier fokuserer dette speciale på Quality of Experience (QoE) indikatorer fra selve videoen: billedopløsning og afspilningsfrysninger. Formålet var at forudsige QoE for ultralav-latens livevideo (ULLV) ud fra målinger i et 4G LTE-netværk. Vi indsamlede data i et levende LTE-net med to telefoner, analyserede resultaterne og brugte et automatiseret klassifikationsværktøj til at træne prædiktionsmodeller. Der blev udviklet to modeller, en til opløsning og en til frysninger, og deres output blev kortlagt til en simpel QoE-indikator fra 0 til 10 for at efterligne en subjektiv Mean Option Score. Modellerne var ikke egnede til scheduling/ressourceplanlægning i LTE, men de kan estimere en QoE-score til overvågning af ULLV. Sammenlignet med de observerede opløsninger og frysninger opnåede forudsigelsen en rod-middelkvadratfejl (RMSE) på 1,63.
Mobile network operators need measures that reflect what users actually perceive. Instead of relying only on network-centric Quality of Service (QoS), this thesis focuses on Quality of Experience (QoE) indicators from the video itself: picture resolution and playback freezes. The goal was to predict QoE for ultra-low-latency live video (ULLV) using measurements from a 4G LTE network. We collected data on a live LTE network with two phones, analyzed the results, and used an automated classification tool to train prediction models. Two models were developed, one for resolution and one for freezes, and their outputs were mapped to a simple 0-10 QoE indicator to mimic a subjective Mean Option Score. The models were not suitable for LTE scheduling decisions, but they can estimate a monitoring QoE score for ULLV. Compared with observed resolutions and freezes, the prediction achieved a root mean square error (RMSE) of 1.63.
[This abstract was generated with the help of AI]
Keywords
Documents