Estimation of the QoE for video streaming services based on facial expressions and gaze direction

As the multimedia technologies evolve, the need to control their quality becomes even more important making the Quality of Experience (QoE) measurements a key priority. Machine Learning (ML) can support this task providing models to analyse the information extracted by the multimedia. It is possible to divide the ML models applications in the following categories: 1) QoE modelling: ML is used to define QoE models which provide an output (e.g., perceived QoE score) for any given input (e.g., QoE influence factor). 2) QoE monitoring in case of encrypted traffic: ML is used to analyze passive traffic monitored data to obtain insight into degradations perceived by end-users. 3) Big data analytics: ML is used for the extraction of meaningful and useful information from the collected data, which can further be converted to actionable knowledge and utilized in managing QoE. The QoE estimation quality task can be carried out by using two approaches: the objective approach and subjective one. As the two names highlight, they are referred to the pieces of information that the model analyses. The objective approach analyses the objective features extracted by the network connection and by the used media. As objective parameters, the state-of-the-art shows different approaches that use also the features extracted by human behaviour. The subjective approach instead, comes as a result of the rating approach, where the participants were asked to rate the perceived quality using different scales. This approach had the problem of being a time-consuming approach and for this reason not all the users agree to compile the questionnaire. Thus the direct evolution of this approach is the ML model adoption. A model can substitute the questionnaire and evaluate the QoE, depending on the data that analyses. By modelling the human response to the perceived quality on multimedia, QoE researchers found that the parameters extracted from the users could be different, like Electroencephalogram (EEG), Electrocardiogram (ECG), waves of the brain. The main problem with these techniques is the hardware. In fact, the user must wear electrodes in case of ECG and EEG, and also if the obtained results from these methods are relevant, their usage in a real context could be not feasible. For this reason, my studies have been focused on the developing of a Machine Learning framework completely unobtrusively based on the Facial reactions.