Football, one of the most popular sports in the world, has 22 players battling with each other for possession of the ball.
In addition to the experience of watching football games, we can learn a great deal from the information we gather from them.
In an attempt to contribute to this subproblem of football analysis, I share how I could not interpret matches using television-like video streams when evaluating matches’ results. To learn more, visit baanstepball.com.
Some aspect of your life bothers you
Although fixed cameras can be placed throughout the field, moving cameras present challenges to obtaining accurate positional and semantic information. Real stadiums, however, would not be able to achieve this, due to their budget and permission constraints. You can process video data in a number of ways without leaving your chair when your budget is tight.
Take a look at what you should do
To manage this big task, we decided to break it down into smaller pieces rather than attempt to break it up into more manageable, more specific chunks like textbook programmers would normally do.
Following are the divisions that we have established in response:
- In a two-dimensional space, the positions of players can be projected through the camera view.
- There should be identification of players, balls, and officials (e.g., their country of origin).
- Object tracking (also known as entity tracking) is essential in my project.
- Is it possible to identify players during a break between frames? Are they identifiable in any way?
- Playing for a team.
We will next examine a particular task, like positioning and semantic data, in greater detail.
During each frame sequence, there is an object detection as a field and an entity detection (field detection). Each entity is tracked when almost consecutive events occur.
In order to estimate the field’s position in relation to the camera, each entity in the field is projected onto the camera. By identifying and placing each player within a team we can also track their performance.
If the video has ended, then the video should be repeated frame by frame until the end is reached. Smoothing occurs when the video ends. By comparing the trajectory paths detected over the sequence with their similarity, we ‘backward adjust’ the data to smooth it out.
You can see what steps are taken immediately after feeding a frame into the system.
Detecting objects using a method
As you work with machine learning, you will quickly notice how difficult it is to find good labeled data. LoV3 is one of the most popular methods for finding objects.
Pre-trained nets and cutting the frame aren’t the best options. Since accuracy is more important than speed, YOLO was used to transmit the original resolution image. When the ball is near a player or referee, they can use this method.
