Understanding Humans in Videos: From Pose to Action Recognition

This doctoral dissertation focuses on designing, developing, and evaluating methodologies for representing and understanding human motion and humans in the scene (interactions with objects or in small groups). The motivation arises from the growing interest in vision-based solutions which can understand and anticipate human behaviours, not only for robotics but also for surveillance or rehabilitation applications. Associated with specific objectives of this thesis, three distinct methodologies are presented: \textit{MOSAIC} which focuses on hierarchical motion representation for action recognition; \textit{ACROSS} which leverages the complexity of activity recognition using scene-graphs representations; \textit{HHP-Net} which is a Head Pose Estimation Network with a focus on interpretability and computational costs. Experiments and discussions support all the presented methods to highlight their strengths and weaknesses and are assessed using state-of-the-art benchmark datasets.