Abstract
In contact sports like rugby union, the tackle is the primary cause of injury, particularly concussions. This paper presents a computer vision pipeline designed to automatically assess injury risk by identifying the spatial relationship between players' heads at the point of contact. By integrating YOLOv4 object detection, Kalman filtering for ball tracking, and OpenPose for skeletal mapping, the system identifies the "tackle frame" and classifies events as high- or low-risk. Evaluation on professional match footage demonstrates a 62.5% accuracy with high recall for dangerous tackles, offering a path toward objective, automated safety monitoring.
Introduction
Enforcing safety regulations in high-speed contact sports is notoriously difficult for referees, whose decisions are often subjective and influenced by external factors. Currently, injury analysis relies on manual video review, which is slow and prone to human bias. This research aims to bridge this gap by creating an objective, automated system that identifies high-risk tackle techniques—specifically those where the tackler fails to drop their body height, leading to dangerous head-to-head proximity.
Methodology
The proposed architecture processes in-game video footage through a multi-stage computer vision pipeline:
- Detection: Two separate YOLOv4 models are used to detect the rugby ball and the players.
- Tracking & Refinement: A Kalman Filter (KF) tracks the ball's trajectory to smooth out noisy detections and handle frames where the ball is obscured.
- Identification: The system identifies the ball-carrier (the player closest to the ball) and the tackler (the player whose bounding box overlaps with the carrier just before contact).
- Pose Estimation: OpenPose maps the skeletal structure of both players, calculating a "head-centre" by averaging facial landmarks.
- Risk Assessment: The "tackle frame" (pre-contact) is analyzed. A high-risk tackle is defined as one where the head-centres of both players align within a specific vertical threshold, indicating an upright tackler.
Datasets and Evaluation
The YOLO ball detector was trained on 551 images scraped from open-source repositories. For evaluation, the researchers used 109 tackle video segments provided by a professional rugby analyst, which served as the ground truth for high- and low-risk classifications.
Key Insights: Head Proximity as a Proxy for Risk
The core innovation of this system is the simplification of tackle risk into the spatial relationship between two head-centres. By focusing on whether the tackler is "upright" versus "low," the system mirrors the technical cues coaches use to teach safe tackling. The system programmatically identifies the exact moment of "pre-contact" based on bounding box overlaps, which is the critical window for injury mechanics.
Results
The system was evaluated on 64 successfully processed tackles from the initial set of 109:
- Accuracy: The system achieved a 62.50% evaluation accuracy in classifying tackle risk.
- Optimal Threshold: A 15% vertical height threshold provided the best balance, yielding the highest F1 Score (0.50) and a Cohen’s Kappa of 0.28.
- Recall: The system showed a high recall for high-risk tackles (up to 82.35%), meaning it is effective at "catching" dangerous tackles, even if it occasionally flags safe ones.
Limitations and Future Work
The primary hurdle for this system is occlusion. When two players collide, their bodies overlap, often causing pose estimation models to lose track of key joints or merge skeletons. This explains why the system could only successfully evaluate 64 out of 109 tackles.
Future directions include moving toward multi-person pose estimation frameworks designed to disentangle overlapping figures and utilizing Recurrent Neural Networks (RNNs) to analyze the entire movement sequence leading up to the hit rather than a single frame.
Summary updated