Tracking and studying players behaviors and statistics has always been paramount in professional sports. Athletes are constantly monitored during training in order to boost their performance by means of ad-hoc sessions and strategies. Games and matches are always recorded by trainers to be further analyzed to extract useful statistics and study both own players and the adversaries. Unfortunately, great part of this useful analysis is customary carried out by expert operators that manually annotate videos and recordings. Thanks to the great advances in computer vision and deep learning, it is nowadays possible to help manual annotators by means of automatic or semi-automatic multimedia analysis techniques. In this work, we propose a solution for the analysis of Badminton players behavior through audio and video analysis. Given a Badminton match video under analysis, we propose a series of techniques that allow: (i) to detect active scenes of interest by means of video frames comparison; (ii) to detect and track players position on the field by means of a Convolutional Neural Network (CNN) adapted to our task; (iii) to detect and track shuttlecock position across frames exploiting an ad-hoc CNN, and; (iv) to detect shuttlecock shots by both players with a novel audio-visual analysis method. Results obtained on a series of annotated videos show the performance achieved by each piece of our system, showcase a preliminary further application of our analysis, and highlight critical issues for future studies.
© Copyright 2022 POLITesi - Digital archive of PhD and post graduate theses. Veröffentlicht von POLITECNICO MILANO 1863. Alle Rechte vorbehalten.