.

Languages and libraries used: -Python -OpenCV -scikit-learn -Flask -HTML,CSS and JavaScript.


This little project will detect offsides in an image input by the user

Here is an example of an input image:

football match

For the moment, after the detection we get the detected players on the image:

detected players
Detected players
After removing players that are not on the field (managers, referees, etc...), we get this bird eye view of the field with the players positions, and the lines that represent the offside positions in both sides:
positions on the field
Positions on the field
We now can go back to the original image and draw the offside lines:
offside lines
"Offside lines" in the origin image

Goal

My idea is to work on a still image of a TV broadcast of a football match following these steps:

  • calibrate the camera to the field to get 3d world coordinate
  • get position of the players
  • identify the players of different teams
  • check if a player is on an offside position

I will discuss each step separately

Type of photos I will be working on

The input images should be a single frame from a footbal game broadcast (Like this). The bigger the field of view the more accurate the detection will be.

One of the features that can be added is possibility of using a video, as such that the offside lines will be drawn in real time

Calibration of the camera

On a single frame, this step is quite easy, the user will need to click on 4 points on the field in the input image, and then click on the corresponding points in a field map. The program will then calculate the homography matrix, which will be used to get the 3d world coordinates of the players.

Get position of the player

Usually, players have at least one foot on the ground, So, their positions on the field is the postion of the feet. There are 2 ways to get the position of a player:

  1. Manually click the feets of all the players
  2. Automatically getting the feet positions(what I wanted to achieve but a bit tricky)
  3. After using OpenCV functions, and a trained machine learning model, I've got the positions of the players.

Identification of the players teams

This is the hardest step. I was thinking about some sort of clustering, then classification. But What if the user had the possibility to pick the color of each team jersey and we then get some algorithm to get the players. Or maybe just enter it manually. I'm not sure about the latter one. I can also look for some sort of deep learning model to get this step right.but it's gonna be the hardest way.

Ball position

The position of the ball is very important, because offside rules depends a lot on it (if it was the last after every attacker)

Check for offside

After all these steps, this gonna be the easiest one, we will just compare the minimum x coordinate of each team. For now, I will not study the edges cases, like goalkeeper before last defender, or ball in other center of half of the field.

To learn more about the offside rules, you can check this.