Implement filter

[amonras]

The main contribution of this PR is the addition of two classes:

• Court3DModel is a class that allows to map 3d world coordinates to 2d coordinates in the image space.
• KalmanFilter3DTracking used to track in 3d the ball location.
• Some unit tests for the new classes, plus some test resources

The Kalman filter also provides a method to estimate the initial location of a parabolic path, given its 2d image. Ideally, this estimate should be used upon initialization, but I have not addressed this for now.

Some additional changes:

• Pin versions in requirements.txt
• Added some files to .gitignore
• The Court model requires 14 keypoints to operate. With less, it will not work. Exception handling for this case has not been implemented.

[RWezeman]

Cool work. 3D detection seems not yet optimal but might have to play a bit more around with it.

Some suggestions I had to try to get your code to work

[amonras]

thanks for your suggestions. Accepted all of them!