PADELVIC: Multicamera videos and motion capture data in padel matches
Recent advances in computer vision and deep learning techniques have opened new possibilities regarding the automatic labeling of sport videos. However, an essen-tial requirement for supervised techniques is the availability of accurately labeled training datasets. In this paper we present PadelVic,...
| Autores: | , , , , , , , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2024 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/400742 |
| Acceso en línea: | https://hdl.handle.net/2117/400742 https://dx.doi.org/10.17398/2952-2218.2.89 |
| Access Level: | acceso abierto |
| Palabra clave: | Deep learning (Machine learning) Computer vision Sports sciences Paddle tennis Pose estimation Player tracking Sport analytic Visión por computador Estimación de la pose Seguimiento de jugadores Análisis deportivo Aprenentatge profund Visió per ordinador Ciències de l'esport Pàdel Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Processament del senyal::Processament de la imatge i del senyal vídeo Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic |
| Sumario: | Recent advances in computer vision and deep learning techniques have opened new possibilities regarding the automatic labeling of sport videos. However, an essen-tial requirement for supervised techniques is the availability of accurately labeled training datasets. In this paper we present PadelVic, an annotated dataset of an ama-teur padel match which consists of multi-view video streams, estimated positional data for all four players within the court (and for one of the players, accurate motion capture data of his body pose), as well as synthetic videos specifically designed to serve as training sets for neural networks estimating positional data from videos. For the recorded data, player positions were estimated by applying a state-of-the-art pose estimation technique to one of the videos, which yields a relatively small positional error (M=16 cm, SD=13 cm). For one of the players, we used a motion capture system providing the orientation of the body parts with an accuracy of 1.5º RMS. The highest accuracy though comes from our synthetic dataset, which provides ground-truth po-sitional and pose data of virtual players animated with the motion capture data. As an example application of the synthetic dataset, we present a system for a more accurate prediction of the center-of-mass of the players projected onto the court plane, from a single-view video of the match. We also discuss how to exploit per-frame positional data of the players for tasks such as synergy analysis, collective tactical analysis, and player profile generation. |
|---|