NFL Health & Safety - Helmet Assignment

Through a collaboration of the NFL(National Football League) and AWS(Amazon Web Services), the Helmet Assignment competition aims for the best sports injury surveillance and mitigation program. As the NFL is trying to gather more data and a sample of exposures for each player in order to estimate the impact rate, the job of tracking each player on a field with 21 different other players clashing with each other is a really difficult task of its own. Thus, the main goal of this competition is to identify and assign football players’ helmets from the given footage, which will subsequently be used to determine the players’ movement on the field. In its simplest form, the goal is to provide real-time object detection in motion. As to provide some insight on what’s the expected performance of a submitted algorithm, the competition describes 90% accuracy as a successful submission.

An example algorithm, taken from competition’s page on Kaggle

An example algorithm, taken from competition’s page on Kaggle

There are a handful of challenges through the development of the algorithm such as blurry frames, variations in different viewpoints/different camera angles, players on the sideline, etc., that our team intends to overcome. There are few built-in resources such as Detectron2 and YOLOv5 that might help the competitors during this competition, but it’s still up to the teams to enhance and adapt the pre-existing structures to the challenge. One of the biggest constraints in the competition is the fact that it almost takes 8 hours to test the algorithm on full data, which is due to the size of the data set. So it is more important to provide complete/working code compared to other competitions as the time to change the existing code is limited. Nonetheless, our team is excited and eager to show what they are capable of!

Previous
Previous

Deep Chimpact: Depth Estimation for Wildlife Conservation

Next
Next

Churn Prediction