Ultimetrics

Using Computer Vision to analyze Ultimate Frisbee Footage.

Project Overview

Between the rise of computer vision technology in sports and the lack of resources for Ultimate Frisbee, we used the OpenCV library to create a tool that would be useful for frisbee coaches to keep track of analytics. Our ultimate objective was to generate a 2D animation that depicts the positional placement of each player on the field based on the drone footage. To do this, we take our film and find the players there using various methods such as detection, tracking, and redetection. Once we have found each of our players, we put a bounding box around them for each frame in our film. We then move on to finding the frisbee field, so we can find the location of each of these players relative to a frisbee field rather than just relative to our field. Once we have completed our perspective transformation, we can keep track of each player's coordinates and produce a 2D animation with the Python library matplotlib.

Terminology

DetectionHow do we locate where players are in the film? We used a machine learning model that uses the YOLO algorithm to find them.
TrackingHow are we following the players on the field? Once we know where the players are, we can use a tracking algorithm that will follow the movement of the players.
RedetectionWhat if we lose a player? Players move around a lot and we can mix tracking up or lose players. We can reuse the detection to locate again where the players are.

Perspective Transformation

How do we translate the field to a perfect rectangle? Using the coordinates on the field and some math, we can translate whatever quadrilateral the field is on the film into a easy to view rectangle.

2D Animation

How do we view all this? Using matplotlib, we can translate this data into a easy to view field that shows us where the players are on the rectangular field we made.

Demo

Here is a demonstration on how our deliverable works.

Input Video

Identify Field Boundaries

Input the Number of Players

Object Detection

Draw Box around Missing Players

Identify Teams

Tracking

Redetection when Players are Lost

Animation

Future Steps

Here are some areas where future groups could build off of our project.

Disc Tracking

Figure out how to track the disc, which has proven to be a challenge due to its small size, unique shape, and rapid velocity.

Improved Detection Model

Use a better dataset so objects can be identified more specifically and teams can be identified based on color.

More Metrics

Look at patterns in player positions, check if certain positions are easier to score from, and possibly measure team chemistry.

More Flexible Program

Allow the program to take in drone footage that does not contain all four corners of the field.

About Us

We are a team of 6 CS student-athletes led by Eric Alexander at Carleton College in the class of 2024. With our shared love of sports and computer science, our goal was to create a visual tool to analyze Ultimate Frisbee Footage.