This project combines 3 of my interests and values in life: Data Analytics, Hockey, and Family. Using YOLOv4 and Python I was able to create a computer vision model which detects, correctly identifies, and tracks individual hockey players, goalies, and the referees in the TV footage of a live game. In doing so, I also created a new tool which takes the output of the detection model to continually create new training data.

What makes this project special is that the video that I trained and tested my model on was the first game my Grandpa ever played in the NHL. On February 3rd, 1962, he entered as an Emergency Back-up Goalie (a la David Ayres) for the New York Rangers in a game against the Toronto Maple Leafs.

Following the creation of my first model, I took what I learned and used it to create a new model on a modern game. I used the ads on the ice to create anchor points and used those known locations in combination with the detected player locations to map their positions.

I aim to continue this project using this research paper to make the model better and be able to create new, meaningful hockey data from broadcast data.

The Github Repository for the training data and the custom Python scripts can be found here.