System for interactive sports analytics using multi-template alignment and discriminative clustering

1. A computer-implemented method for interacting with a user via a graphical computer interface to interactively analyze a play of a sports game, the sports game having been tracked with a tracking system to record two-dimensional positional information for objects on the playing surface and event data having been determined for the game and associated with the two-dimensional positional information, and the play having been retrieved from a database of historical plays and event information and loaded onto the interface, the method comprising:

displaying a graphical depiction of the play on the interface, the graphical depiction representing the play retrieved from the database and including a representation of the playing surface and including markers representing positions of tracked objects at a given point in time;

presenting a statistical value for an aspect of the play at the given point in time, the statistical value generated from the database of historical plays and event information;

receiving a proposed modification to the graphical depiction of the play in the interface, via user interaction with at least one marker;

querying the database with the proposed modified play;

obtaining from the database a statistical value for the aspect of the proposed modified play, the updated statistical value being generated based on the historical plays and the game event information in the database; and

updating the presented statistical value with the obtained value for the aspect of the proposed modified play.

2. The method of claim 1 , wherein the one marker represents a first player, and the proposed modification is a substitution of a second player for the first player.

3. The method of claim 1 , wherein the one marker represents a ball, and the proposed modification is a relocation of the ball.

4. The method of claim 1 further comprising:

presenting an animation of the graphical depiction of the play for a window of time according to the tracked positional information, the animation including a depiction of object trajectories over the time window; and

displaying a user-selectable interface option for refinement of the time window;

wherein the proposed modification is a change in the trajectory of an object during the time window.

5. The method of claim 1 further comprising:

obtaining from the database a ranked listing of historical plays according to a calculated similarity to the queried play, the similarity based on at least one of: positional information; event information; contextual information; and

presenting a graphical depiction of one or more of the most similar historical plays on the interface.

6. The method of claim 5 further comprising:

obtaining real-world video corresponding to the presented most similar historical plays; and

playing the video via the interface.

7. The method of claim 1 , wherein the given sport is basketball and the statistical value is an expected point value for a player.

8. The method of claim 1 , wherein the given sport is soccer and the statistical value is an expected goal value.

9. A method for indexing a database of plays for a sport, each play including two-dimensional positional tracking data for the objects on a playing surface and a time interval, the method comprising:

selecting a first play for use as a first template;

determining a set of plays for use as additional templates;

aligning the plays in the database to each of the templates, representing each aligned play in the database with a permutation matrix for the play; and

discriminatively clustering the aligned plays in a top-down hierarchical manner to form a hash table/decision tree;

wherein determining the additional templates comprises:

calculating a cost matrix for each play in the database with respect to the first template, the cost matrix representing a quantitative difference between the trajectories of each tracked object in the play and each object in the template;

determining a permutation matrix for the play with respect to the first template;

choosing a value K>1;

clustering the plays in the database to assign them into K clusters around K candidate plays;

calculating the total reconstruction error for the K clusters, and using the K candidate plays as the additional templates if the total reconstruction error is less than a desired threshold value.

10. The method of claim 9 , wherein discriminatively clustering the aligned plays comprises:

assigning each aligned play to one of the K clusters using the two-dimensional tracking data for the objects in the play to form a first level of the hash table/decision tree;

calculating a value of a classification metric for each play in the first-level clusters;

applying a back-propagation technique to each play in the first-level clusters to determine a subset of objects whose trajectory information contributes significantly to the value of the classification metric, and adjusting weights for the objects classification metric correspondingly; and

assigning each aligned play of a first-level cluster to a second-level sub-cluster, based on application of the re-weighted classification metric.

11. The method of claim 10 , wherein the sport is one of: basketball, soccer, rugby, American football or volleyball.