Final Results
Ultimetrics
Advisor: Eric Alexander
Background
For many years, professional sports teams have made use of increasingly advanced and fine-grained metrics to evaluate players and tactics. Long gone are the days of making sweeping judgments based off of the broad aggregates contained in box scores. Now, leagues and news outlets alike are able to construct models built on the precise locations and behaviors of every player at every second of every game. Having access to such detail has often required hours upon hours of dedicated work of many people meticulously annotating game footage, which ends up being very expensive. However, increasing availability of high quality cameras along with advances in computer vision have begun to make such techniques available to coaches, teams, and sports that operate on much smaller budgets.
One such sport is ultimate frisbee. Ultimate is a possession-based team sport with similarities to soccer, basketball, football, and more, though it has not been around nearly as long as these others. Up until very recently, most analysis in the sport has relied entirely on qualitative judgements and the tallying of discrete events (e.g., goals, throws, drops, etc.). As high quality game footage becomes more and more common, however, coaches have begun to think about how they might use positional data to answer more subtle questions. These include things such as how well players are spacing themselves on the field, whether certain players “draw” defenders in a way that creates opportunities for their teammates, or if there are identifiable clusters of opponents’ play styles that can be anticipated and taken advantage of. Recording such data purely by hand is prohibitively expensive, and it is unlikely to be completely automatable, but a combination of computer vision and human annotation has the potential to dramatically expand the type of analysis coaches are able to perform.
The project
In this project, you will be working with the coaches of Carleton ultimate frisbee teams to develop a tool to help them annotate game footage, automatically extract advanced metrics, and visualize these metrics in ways that support their decision making. This will require you to:
- Perform extensive contextual inquiry and user modeling to learn the needs and desires of the coaches.
- Investigate advanced positional metrics used in other sports and work with the coaches to create new metrics for ultimate frisbee.
- Leverage existing libraries like OpenCV to automate object (player) detection and tracking within game film where possible.
- Build an interface designed to streamline input from a human annotator when needed to train the CV model and disambiguate when discrepancies arise.
- Construct interactive visualizations of the positional metrics that inform both player specific and teamwide analysis.
- Meet consistently with the coaches to perform iterative evaluation and development of your tool(s).
Deliverables
A human-in-the-loop tool for analyzing game footage for ultimate frisbee, along with associated documentation.
Recommended experience
Ideally, some members from the team will bring experience from CS 344 (HCI) and/or CS 314 (Data Vis). CS 257 (Software Design) is also likely a plus, as is any experience with computer vision. However, it will not be required that all individuals working on the project have this background.
Note: you do not need to know anything about ultimate frisbee for this project! Interest in sports, computer vision, and application design will help make the experience enjoyable for you, but the project will benefit from a diversity of experiences and background knowledge.