This project was undertaken as part of the final year of my undergraduate degree, starting in Autumn 2011 and finishing in Summer 2012. It focuses on recognising board game structures from images, in such a way that it is easy to recognise new games. The two test games were Connect Four and Noughts and Crosses, and the system I created was tested on 100 images of each game captured by a camera, as well as a large number of images generated artificially.
The dissertation provides detailed results, but I’ll summarise here: all images captured by the camera were recognised correctly, and images of both games were recognised at large distances, at low lighting and at significant pitch rotations (and all roll rotations). Performance was also good: Connect Four boards were recognised on average in about 9 milliseconds, while Noughts and Crosses boards were recognised on average in about 3 milliseconds.
Here is the dissertation. Code is available at the repository.