Player motion analysis : automatically classifying NBA plays

Abstract

Basketball is a team game, and an important task for coaches is analyzing the effectiveness of various offensive plays. Currently, teams spend a great deal of time examining video of past games. If teams could automatically classify plays, they could more effectively analyze their own plays and scout their opponents. In this thesis, we develop a methodology to help automatically classify a set of NBA plays using data from the SportVU optical tracking system, which tracks the position of each player and the ball 25 times per second. The problem is made challenging by the variations in how a play is run, the high proportion of possessions where no set play is run, the variance in length of plays, and the difficulty of acquiring a large number of labeled plays. We develop a framework for classifying plays using supervised machine learning. In our approach, we incorporate a novel sliding block algorithm that improves our classifier by accounting for the difference in play lengths. We also use a variant of the traditional one vs. all multi-class SVM. This approach is well suited to distinguish labeled plays from free motion and unlabeled plays. This thesis demonstrates that we can use SportVU data to automatically differentiate plays. We selected a total of six plays to classify, where each play had at least 20 labeled instances. We also added a large selection of plays that were not one of these six and labeled them as Other. Our framework correctly predicted the play with an accuracy of 72.6% and an F-score of .727. We also propose a framework, based on our engineered features, to extend our research to unlabeled plays.