Classification of passes in football matches using spatiotemporal data

A knowledgeable observer of a game of football (soccer) can make a subjective evaluation of the quality of passes made between players during the game, such as rating them as Good, OK, or Bad. In this article, we consider the problem of producing an automated system to make the same evaluation of passes and present a model to solve this problem. Recently, many professional football leagues have installed object tracking systems in their stadiums that generate high-resolution and high-frequency spatiotemporal trajectories of the players and the ball. Beginning with the thesis that much of the information required to make the pass ratings is available in the trajectory signal, we further postulated that using complex data structures derived from computational geometry would enable domain football knowledge to be included in the model by computing metric variables in a principled and efficient manner. We designed a model that computes a vector of predictor variables for each pass made and uses machine learning techniques to determine a classification function that can accurately rate passes based only on the predictor variable vector. Experimental results show that the learned classification functions can rate passes with 90.2% accuracy. The agreement between the classifier ratings and the ratings made by a human observer is comparable to the agreement between the ratings made by human observers, and suggests that significantly higher accuracy is unlikely to be achieved. Furthermore, we show that the predictor variables computed using methods from computational geometry are among the most important to the learned classifiers.