It looks like it was just a bunch of video from Bundesliga, and only had frames marked in terms of the type of play that the clip was part of. Briefly digging he seems to have done a fine-tuning of whatever YOLO version he was doing with a hand-made dataset of just 300 images (split between train/validation/test) at https://universe.roboflow.com/roboflow-jvuqo/football-players-detection-3zvbc/dataset/12.

The video he is showing is repurposing that video to fine-tune YOLO to detect and track players, ball, and ref, with the final results being a summing of the amount of travel made by a player. YOLO is pretty good, but would work better with the fine-tuning for just soccer players.

The guy does provide his trained model, but not his pre-processed video data with the jersey extraction, bounding boxes, etc. You would have to re-run that part of his analysis and it looks to be done just by OpenCV k-means.

I think you could run-it as-is without the YOLO fine-tuning, but he claims he was doing a re-train.