Automatic editing of football highlights

Football events are manually tagged by observers watching a game, however the starting point and end point of a highlight clip is typically not. Develop a machine learning algorithm to optimize the span of the video clips.
Master

The student will analyze video clips using various machine learning algorithms to provide video highlights that are interesting for supporters to watch. The system is currently serving the Norwegian and Swedish top football leagues, and training data for professionally created highlights is provided as training sets. A trivial algorithm baseline subtracts a fixed number of seconds from an event to mark the start, but the student should aim to improve this by detecting the length of the interesting section leading up to the event by analyzing the video and the players positional data, when available.

Goal

Develop and evaluate a machine learning system for detecting span of football highlight events. Academic publication of results.

Learning outcome

Practical experience in using machine learning techniques for real video events.

Qualifications

Understanding of neural nets and other machine learning techniques. INF5860 recommended.

Supervisors

Collaboration partners

ForzaSys AS