Resumen de Contributions to head pose estimation methods
Head Pose Estimation (HPE) is currently a growing research field, mainly because of the proliferation of human-computer interfaces (HCI) in the last decade. It offers a wide variety of applications, including driver assistance systems, pose-invariant face recognition, human behavior analysis, or popular HCI applications such as gaze estimation systems. HCIs show an increasing tendency to integrate HPE as another bridge for interaction, since it is a rich form of communication. For instance, gaze tracking systems suffer in unconstrained environments because they are very sensitive to head motion, and HPE has become a key point for successful gaze estimation. This thesis thus aims to contribute to the development of robust and accurate HPE methods based on 2D tracking of the face in videos.
With the idea of achieving a better understanding of every aspect of the HPE process, a complete framework has been created in the first part of the thesis as a pillar to sustain the rest of the work. This framework consists of both simulation and realistic environments for HPE algorithm analysis. It includes the recording of two head pose databases of videos, one with synthetically generated heads and the other one with real subjects. They have proven to be extremely useful tools for the purpose, and therefore we expect to make them available for the whole scientific community.
The problem of 3D face reconstruction using only 2D images from the videos has received special attention. A whole chapter has been devoted to the study and comparison of different single-view and multi-view based reconstruction methods in a controlled simulation environment. This has allowed us to isolate the 3D model fitting problem, thus drawing several conclusions regarding the influence of this critical part in a HPE system.
With the aim of achieving a wider impact with this thesis, the pose estimation problem is addressed from a general perspective in which techniques that are generalizable to any kind of 3D object are proposed. Starting from a basic pose estimation approach (2D tracking & POSIT), different alternatives have been developed to improve performance. On the one hand, a tracking accuracy index (TAI) calculation method has been proposed, based on invariant shape metrics obtained from interlandmark relationships. This allows us to apply weights that compensate for 2D tracking inaccuracies and optimize the 3D pose estimation. On the other hand, outlier detection and outlier correction methods that aim to improve the 2D tracking itself have been proposed, addressing the typical drifting problem of point-tracking systems, and hence improving the 3D pose estimation further. These global methods have then been specifically adapted to HPE and evaluated using two head pose databases: our real database, which reflects the expected performance in current technological conditions, and the BU database, a widely referenced older database that allows an extensive comparison with other state-of-the-art HPE methods.
