HN Books @HNBooksMonth

"It's tracking based on hardcoded colours. Try this on a video with even the slightest colour variation, and it will fall apart."

How would you improve it? (Serious question.) Some sort of training data set of many different footballs from many different games? It's easy to offer criticism without also offering how this could improve.

I personally think this kind of thing could be the future of television - heavy in-game analysis for sports stat nerds.

At the moment football (i.e. soccer) teams pay a lot of money to get stats on their own players[1] (distance ran, passes complete, assists, etc, etc). This kind of thing could lower the cost of gathering this data. I'm not an expert in the area, but I read somewhere that at the moment at least some stats are gathered manually (guy sitting in the stands.) It would be of particular interest to clubs doing the whole soccernomics thing (idea adopted from "Moneyball" in US baseball). Some clubs practicing soccernomics: Liverpool in England and Lyon in France [2]. Could be a serious startup opportunity here!

[1] http://www.optasports.com/sports/football.html

[2] See this book for more: http://www.amazon.co.uk/Soccernomics-England-Germany-Austral...

⬐ andypants

Well, thresholding based on colours is really as basic as it gets. Of course, a lot of the problems come from the quality of the video you're working with.

As for automated player stat tracking, it's probably not easy, but I think it is doable. Computer vision and object tracking is definitely already being applied to american football, although I'm not sure if it's being applied to the ball and players (it's being applied to the field and stadium to show scores and yard lines).

Although, there would be a lot of huge problems, such as video feed quality, the video feed itself (you probably don't want one with tv channel overlays, lots of zoomed-in shots, replays, etc., which would make it difficult to find a good feed), and probably more problems which will pop up as you try to code the program. It would probably be useful to have some kind of partnership with whatever authority in charge of match broadcasting.

Examples of improvements: pre-process the image and adjust the colour balance, etc. You can track more than just colour. Relax the colour thresholds, and include thresholds for area, density, etc. You might also want to automatically calculate the threshold values, rather than hand coding them in.

The wiki page for segmentation ( http://en.wikipedia.org/wiki/Segmentation_(image_processing) ) lists plenty of alternatives to colour thresholding too.

Since a video is just a series of images, it would make sense to use previous images to help track anything in a video. A Kalman filter ( http://en.wikipedia.org/wiki/Kalman_filter ) with condensation tracking ( http://en.wikipedia.org/wiki/Condensation_algorithm ) can statistically predict object positions and should also be able to cope with fast changes in direction. Also useful for frames where the ball disappears behind a player or something.

Stuff like image moments ( http://en.wikipedia.org/wiki/Image_moment ) can describe objects despite movement, rotation and scaling, so that would be useful if the camera zooms in/out.

I'm probably also missing out a whole bunch of more advanced vision stuff. I'm no expert, sorry.

I don't think soccernomics would be quite as useful as moneyball though, since players in soccer depend on each other way too much for individual stats to mean much, while in baseball, the players pretty much play as individuals, which is why it's so easy to track stats and compare players.

Edit: Also, to see some examples of the problem I mentioned in my previous post, the video identifies a circle inside the penalty box as the ball. And the ball is actually yellow, but that's only clear when the camera zooms in, so nothing is being detected during those shots.

I feel like I may be getting a bit too critical of the program, so apologies to the author. I really do think it's quite an awesome demo despite everything I mentioned.

⬐ onemoreact

One of the major improvements that human vision includes is predictive decoding. Basically, how you interpret an immage is based on your model what you expected that immage to look like.

In the case of football, you might build a model of probable footballs for each frame then try to connect them frame by frame until you create a reasonable path and then say ok that's where the actual football is. Then for each new frame you keep extending that path based on what's in the frame and how far the ball can reasonably travel.

PS: Classic example, you tend to swap between vase or face, not see them both at the same time. http://en.wikipedia.org/wiki/Rubin_vase Or the more striking http://www.michaelbach.de/ot/fcs_hollow-face/index.html

A fascinating book to read as the World Cup nears:

http://www.amazon.com/Soccernomics-Australia-Turkey-Iraq-Are...