Abstract Master Thesis by Rejnald Lleshi:
Action recognition is a task that has proven central in gaining insights from videos. It has received growing attention in computer vision and has improved significantly in recent years. Its goal is to identify which human actions appear at which point during a video. Video data is particularly challenging to work within computer vision because it contains complex temporal dynamics due to the variety of actions, speed variations, or camera motion. This thesis seeks to apply such methods in a dance-like context for didactical purposes. Specific, dance-like movements from the BAST analysis were studied and evaluated with standard and state-of-the-art algorithms from the field. The BAST analysis examines the relationship between movement and inner psychological state and is typically employed in domains such as psychiatry or clinical psychology. However, because BAST utilizes patterns originating from the Laban analysis, a famous movement analysis methodology in dancing, it could be helpful to investigate and try to extend such movements in a dancing context. Different models are trained on BAST material and evaluated in possible dancing contexts to examine their robustness. Various modalities for feature representation, such as RGB frames, optical flows, and human skeletons, were utilized.