Abstract
The National Football League (NFL) and Amazon Web Services (AWS) collaborated to enhance sports injury surveillance through a Kaggle competition. The primary goal is to assign specific player IDs to detected helmets in game footage, enabling precise tracking of "exposures" or impacts throughout a play. This paper details a computer vision-based machine learning pipeline that integrates object detection, temporal tracking, and geometric alignment to automate the monitoring of player safety.
Introduction
In American football, monitoring head impacts is vital for long-term health and injury prevention. While the NFL utilizes sensor-based tracking data (GPS and accelerometers) for all players, this data lacks direct correspondence to video footage. Identifying which helmet in a 2D video belongs to which player is particularly challenging during high-speed collisions and player clusters. This research addresses the "data fusion" problem: mapping 2D video detections from sideline and endzone cameras to 1D ground-truth tracking data to automate impact surveillance.
Methodology
The proposed system utilizes a multi-stage pipeline to bridge raw video pixels with player identities.
1. Detection and Tracking
The pipeline begins with YOLOv5x6 for robust helmet detection. To optimize performance, the model focuses on frames following the "snap" and employs data augmentation, such as flipping endzone images, to improve generalization across camera angles. Following detection, DeepSort maintains player identities across frames by combining a Kalman filter for motion prediction and the Hungarian algorithm for matching, effectively handling temporary occlusions.
2. Player Assignment and Data Fusion
The core of the system is the alignment of 2D video coordinates with ground-truth tracking data.
- Iterative Closest Point (ICP): This geometric algorithm aligns detected helmet coordinates with the NFL's tracking data.
- Refinement via Clustering: K-means clustering based on helmet color distinguishes between teams, significantly reducing the search space for the assignment algorithm.
- Homography: Transformations are applied to maintain spatial consistency between adjacent frames despite camera movement.
Datasets
The study leveraged the NFL/AWS Kaggle dataset, which includes:
- Video Data: 120 training videos providing both sideline and endzone perspectives.
- Tracking Data: A comprehensive CSV containing x-y coordinates, orientation, speed, and acceleration for every player.
- Ancillary Data: Labeled images with bounding boxes specifically for helmet detection training.
Key Insights and Results
Exploratory analysis revealed that filtering out frames before the "snap" significantly reduces noise, as player movement is minimal during this period. Furthermore, treating player assignment as a geometric alignment problem rather than a pure classification task proved highly effective.
The system achieved a Public Leaderboard score of 0.797 and a Private Leaderboard score of 0.713. While the model demonstrated 89% accuracy for clearly visible helmets, performance decreased for blurred or occluded views typical of high-motion sports.
Conclusion and Future Work
While the current pipeline successfully maps player IDs to helmets, future improvements could focus on direct collision event detection using motion dynamics. Additionally, addressing the slight temporal misalignment (1-7 frames) between different camera views remains a priority for perfecting multi-view tracking. Computational constraints limited the use of more advanced architectures like Detectron2, suggesting that further optimization could yield even higher accuracy.
Summary updated