An Approach to Temporal Phase Classification on Videos of the Volleyball's Basic Reception Technique

No hay miniatura disponible
Brandelli A.
Chávez-Hidalgo E.
Flores-Fernández C.N.
Flores-Santos J.C.
Izaguirre V.
Zavaleta A.I.
Título de la revista
Revista ISSN
Título del volumen
Association for Computing Machinery
Proyectos de investigación
Unidades organizativas
Número de la revista
In this paper we provide an approach on sports analysis using Deep learning techniques. As part of a current project, the volleyball's basic reception technique has been divided into temporal phases. We performed an evaluation over our own labelled dataset consisting in 14814 frames from 69 videos depicting the desired reception technique. A model based on the YOLO algorithm was trained to locate the player region and trim the frames. Two time fusion methods over the frames wereproposed and evaluated with CNN models which were created based on the ResNet models and a transfer learning approach was used to train them. The results show that these models were able of classifying the frames with their corresponding phase with an accuracy of 92.21% in our best model. Also it can be seen that the RGB merging method shown in this paper helps to slightly improve the performance of the models. Furthermore, the models were capable of learning the temporality of the phases as the mistakes done by the models occurred between consecutive phases. © 2020 ACM.
Palabras clave
Volleyball, Activity Recognition, Computer Vision, Sport Analysis