This article describes how the NFL partnered with AWS to use machine learning and AWS Batch to automatically detect helmet impacts in game videos, replacing manual annotation.

• NFL needed comprehensive helmet impact data across multiple seasons for player safety initiatives
• Manual annotation was expensive, slow (1 hour per play), and inconsistent across annotators
• Solution uses three ML tasks: snap detection, helmet detection/player assignment, impact classification
• AWS Batch configured with P3 and G4dn GPU instances across three availability zones
• AWS Step Functions orchestrates workflows with nested state machines for modularity and scalability
• Caching logic skips redundant processing, saving over $1,000 per hour in compute costs
• Results: $700K/season savings, 90% reduction in manual labor, 12% better accuracy than humans
• Compressed 24 years of computation into less than 6 weeks using parallel processing

The NFL now has the first comprehensive historical dataset of helmet impacts since 2016, enabling data-driven decisions on equipment, rules, and coaching to improve player safety.