Dynamic head pose estimation in varied conditions using Dlib and MediaPipe

Abstract

This paper presents the formulation and validation of a dynamic head pose estimation (HPE) algorithm, addressing challenges related to diverse conditions, complex poses, and partial obstructions. The study aims to create a robust algorithm that maintains high accuracy in real-time applications across varying conditions. The algorithm was implemented and assessed using Dlib and MediaPipe models. The study involved 30 participants in face and head without obstacles, face with obstacles and head with obstacles conditions. The results demonstrated impressive performance in both controlled and spontaneous head movement categories. The algorithm achieved an average accuracy of 93% for head pose estimation and 88% in detecting visual attention under spontaneous head movement categories. A correlation coefficient of 0.866 indicates a strong positive linear association between performance and attention accuracy, indicating that performance improvements are intricately linked to proportional increases in attention accuracy. However, this does not necessarily imply causation. The findings provide valuable insights into the effectiveness of the proposed algorithms in assessing visual attention and demonstrate their potential applications in healthcare monitoring, educational intervention, and driver monitoring systems. The significance of these results lies in the ability to advance human-computer interaction, enhance healthcare diagnostics, and offer innovative solutions across various domains.