With the introduction of concepts for virtual interaction and digital doubles, a rich scenario has been created for embodied avatars to strive. These avatars, more recently referred to as digital humans, have become a popular area of research, resulting in various techniques and methods that focus on improving the perception of their realism, fidelity, emphatic response, and interactivity. This survey aims to explore the literature and recent advancements on the key processes behind the creation and animation of digital human faces through the view of a general pipeline. The extensive review carried out in this study explores the usual data collection protocols, the main facial codification paradigms and databases, the approaches for digital human asset creation, facial tracking solutions for performance-driven animation, the solving process, and the final rendering delivery. Different quantitative evaluation methods, visual perception tests, and empathetic response evaluations for digital humans are also included in the survey. Additionally, the paper presents an updated summary of public and private frameworks for digital humans that go through the complete general pipeline presented. Finally, the condensed knowledge is discussed, inquiring into the possible direction of future developments in the field.

Due to the increasing demand for virtual avatars, there has been a recent growth in the research and development of frameworks for realistic digital humans, which create a demand for realistic and adaptable facial motion capture systems. Most frameworks belong to private companies or represent high investments, which is why the creation of democratized solutions is relevant for the growth of digital human content creation. This research work proposes a facial motion capture framework for digital humans with the use of machine learning for facial codification intensity regression. The main focus is to use coded face movement intensities to generate realistic expressions on a digital human. The ablation studies performed on the regression models show that Neural Networks, using Histogram of Oriented Gradients as features, and with person-specific normalization, present overall better performance against other methods in the literature. With an RMSE of 0.052, the proposed framework offers reliable results that can be rendered in the face of a MetaHuman.