Temporal Segmentation of Actions in Fencing Footwork Training

Abstract

Automatic analysis of actions in sports training can provide useful feedback for athletes. Fencing is one of the sports disciplines in which the correct technique for performing actions is very important. For any practical application, temporal segmentation of movement in continuous training is crucial. In this work, we consider detecting and classifying actions in a sequence of fencing footwork exercises. We apply pose estimation to RGB videos and then we perform per-frame motion classification, using both classical machine learning and deep learning methods. Using sequences of frames with the same class we find data segments with specific actions. For evaluation, we provide extended manual labels for a fencing footwork dataset previously used in other works. Results indicate that the proposed methods are effective at detecting four footwork actions, obtaining 0.98 F1 score for recognition of action segments and 0.92 F1 score for per-frame classification. In the evaluation of our approach, we provide also a comparison with other data modalities, including depth-based pose estimation and inertial signals. Finally, we include an example of qualitative analysis of the performance of detected actions, to show how this approach can be used for training support.