End-to-End High-Risk Tackle Detection System for Rugby

Reducing risk of severe injury such as concussion is a high priority for any contact sports. In rugby, Head Injury Assessment (HIA) protocol has been introduced to identify and protect players showing symptoms of concussion and having potential risk of concussion. However, on-field decisions by offi...

Full description

Saved in:
Bibliographic Details
Main Authors: Nonaka, Naoki, Fujihira, Ryo, Nishio, Monami, Murakami, Hidetaka, Tajima, Takuya, Yamada, Mutsuo, Maeda, Akira, Seita, Jun
Format: Conference Proceeding
Language:English
Subjects:
Computer vision
Conferences
Deep learning
Head
Pattern recognition
Pose estimation
Protocols
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Reducing risk of severe injury such as concussion is a high priority for any contact sports. In rugby, Head Injury Assessment (HIA) protocol has been introduced to identify and protect players showing symptoms of concussion and having potential risk of concussion. However, on-field decisions by officials are sometimes difficult and subjective, and HIA is affordable only for elite leagues since it requires medical specialists. To make rugby matches more safe, we aim to develop a system to detect high-risk tackles, potential triggers of concussion, based on deep learning models. Our system takes rugby match video, then first identifies frame with tackle, subsequently detects location of tackle and estimate pose of the ball carrier and the tackler, and finally evaluate the risk of tackle using posture pair of players. Among the model combinations we have examined, the best performance was achieved with the combination of ResNet (2+1)D as tackle frame selection model, RetinaNet as tackle detection model and CenterTrack as pose estimation model. Evaluation using test data, a set of short clips from broadcasted rugby match videos, showed our system was able to detect 50% of high-risk tackles without any human intervention. This result opens a path for automated systems to detect high-risk events, leading to less expensive and more objective monitoring not only for rugby but also for any contact sports.
ISSN:2160-7516
DOI:10.1109/CVPRW56347.2022.00399