TrackNet CNN Architecture for Court Detection of real-time video broadcast 

Demo video showing the results of a convolutional neural network (CNN) based model for doing (near) real-time court detection from a single camera. The model should generalize to arbitrary courts (basketball, tennis, badminton, etc.), but not tested and only trained on tennis-courts as of yet.
The use-cases are mostly analytics / analysis of play, or use as base for ball- and player tracking.
Model is built on the architecture of TrackNet originally used for ball detection by its authors.
The video shows some jittering for some frames, this could be reduced by not computing the homography matrix at every frame but just refining it.
Sources
TrackNet: arxiv.org/abs/1907.03698
Source code:
github.com/SebastianBitsch/te...
Video credits:
• CRAZY 70-Shot Rally! |...
• Incredible 53-Shot Ral...
• Zverev Wins Incredible...
• Djokovic v Federer sen...

Наука