Preliminary Approach to Semi-Automatic Classification of Breaking Movements Using Pose Estimation

Breaking was recently introduced in the Paris 2024 Olympic Games. Among key performance indicators, dancer kinematics should be monitored to analyze the originality and creativity of battles. Currently, the five components of breaking (i.e., Toprock, Footwork, Powermove, Acrobatics and Freeze) are analyzed using notational analysis from video recordings, as in [1], where the authors investigated the temporal and sequential structure of a round. However, this method can be time-consuming and prone to analyst errors. Therefore, the present study aims to apply opensource libraries for full-body pose estimation to distinguish between the five main components of breaking (see Figure). MediaPipe was employed as a computer vision technique to perform full-body pose estimation on video footage of dancers. Since the corresponding landmark model includes 33 keypoints, we refined it to the primary joints (n=7 points) to apply rules similar to [2], facilitating movement classification (Figure). The model was tested on 10 competition round videos previously tagged by an expert video analyst from the breaking federation, with each movement classified as one of the five components, or NaN when unclassifiable (around 10k frames were classified). Basic kinematics information (i.e., 2D coordinates, velocities and accelerations of landmarks) was sufficient to distinguish the main components of breaking.