darwinsbuddy / foosball-ai Goto Github PK

1. dewarp video (based on chessboard or marker based calibration)
2. detect edges of the table
3. resize used frame mask according to table dimensions on dewarped image

and show it in the rendered image :)
potentially after a goal

to live stream per IP or to youtube/social media

Following scenarios should be circumvented which appeared in a test game:

• ball track lost for some frames, reappearing and geting lost again within the goal boundaries
• ball track lost for a couple of frames "near" a goal and counted as goal (as it was not lost factually inside a goal, but within the wrongly recognized boundaries)

Solution:

Make analyzer count a goal only after some grace period of x frames or x fractions of a second without a track re appearing

after the blob detection, in case more than 1 blob has been detected, take also into account

• the blob size (filter out blobs above certain threshold)
• maybe the distance between the last detected position and the current one (filter if blob moved too fast, e.g. false positive drew attention to it on the screen edge)

we are not capturing high speed (native fps)

investigate if we need buffering or maybe another input capture than opencv, or maybe if it's simply an issue of hardware resources

after #3

• refactor code s.t. we can use merely signals of the ball (i.e. position, visibility, last known position) to use it in goal detection
• try to detect goals according to position of the markers on the table or the edges of the goal (dark rectangle?)
• count goals according to signals:
  • last position of the ball within the goal area
  • exceeded threshold time were it hasn't been detected

https://universe.roboflow.com/simon-klemp-rwth-aachen-de/tiny-yolo-ball-detection

As of now is done manually.
Come up with a solution for automatically calibrating it by holding it into the camera (maybe within a colored box?)

implement support for

• ball trajectories
• change in velocity vectors
• player/rod detection

subsume all in origin of shot estimation