Hybrid head pose estimation framework for assessing visual attention in robot–assisted therapy

Visual attention plays a crucial role in learning and therapeutic engagement; however, its assessment among children with special needs remains a challenge due to dynamic, non-verbal behaviours. Existing robot-assisted therapy systems do not offer an automated, reliable mechanism for quantifying attention in real time. To address this gap, this study developed a hybrid Head Pose Estimation (HPE) framework that integrates MediaPipe and Dlib, optimised for mobile robotic platforms to facilitate realtime visual attention monitoring. The study aimed to develop and evaluate a robust, real-time HPE algorithm designed to quantify visual attention in children with special needs during robot-assisted sessions. This algorithm incorporated hybrid facial landmark detection techniques by integrating MediaPipe and Dlib, thereby enhancing accuracy across various head orientations and occlusions. Validation was conducted with 90 participants: 30 adults, 30 typically developing children, and 30 children with special needs.