Here’s what you need to know about football analysis
One of the world’s most popular sports, football, involves 22 players competing for possession of the ball at the same time.
Although watching football games is a part of the experience, we can learn a lot from the data we can collect from them.
As part of my contribution to this subproblem of football analysis, I tell the story of how I was unable to interpret matches using television-like video feeds in order to evaluate the results of the matches. Visit the website baanstepball.com for more information.
You are bothered by something
Moving cameras present a challenge to obtaining positional and semantic information, despite the fact that fixed cameras can be placed throughout the field. The budget and permission constraints of real stadiums, however, would prevent them from achieving that. When your budget is tight and you don’t want to leave your chair, video data can be processed in a variety of ways.
Here’s what you should do
We decided to break up this big task into smaller bits rather than try to break it up into more manageable, more specific chunks as would be the custom for textbook programmers.
We have created the following divisions in response:
- Using the camera view, the positions of players can be projected into two-dimensional space.
- Players, balls, and officials (e.g., where they are from) should be identified.
- In my project, tracking objects (also known as entities) is crucial.
- Do you think you can identify players in between frames? Can you identify them somehow?
- The team that a player plays for (how can I find out).
Taking a specific task, such as positioning and semantic data, as the next step, we will investigate it in more detail.
In each frame sequence, objects are detected as fields and entities are identified (field detection). The entities are tracked when almost consecutive events are detected.
The position of each entity in the field is projected onto the camera in order to estimate the field’s position relative to the camera. We can also track the performance of each player by identifying and placing them within a team.
In order to repeat the video frame by frame until the ending has been reached, once the video has ended. Once the video has ended, the data is smoothed. By looking for similarity in the detection of trajectory paths and trajectory paths over the sequence, we do ‘backward adjustments’ to smooth the data.
Following the feed of a frame into the system, you can immediately see what steps are taken.
A method for detecting objects
One of the first things you notice when working with machine learning is the difficulty in finding good labeled data. One of the most popular object detectors is loV3.
Cutting the frame and using a pre-trained net isn’t the best approach. For transmitting the original resolution image, YOLO was used since accuracy is more important than speed. This method allows players or referees to find out when a ball is near them.