Evaluación de distintos algoritmos de IA para mejora de Head Pose Estimation para sistemas webcam-based eye-tracking

This thesis evaluates multiple AI algorithms to enhance Head Pose Estimation (HPE) in low-cost, webcam-based eye-tracking systems, aiming to reduce errors that affect Gaze Estimation. The study compares state-of-the-art algorithms, including Mediapipe, DECA, and 3DDFA_V2, against a Baseline method, using the UPNA Head Pose Estimation Database. Key metrics analyzed include HPE rotation and translation, 3D eye position accuracy, and inference time. Results show Mediapipe is the best algorithm for real-time HPE due to its balance of accuracy and speed. This work underscores the significance of precise HPE for effective webcam-based eyetracking and suggests further improvements through advanced machine learning techniques. Future research will refine these algorithms, enhancing the accessibility of accurate eye-tracking solutions.