Αναγνώριση ενεργειών σε βίντεο αγώνων ποδοσφαίρου με μηχανική μάθηση
Action detection on football match recordings using machine learning
View/
Keywords
SoccerNet ; Player centered ball action spotting ; Graph neural networks ; Heterogeneous graph transformer ; Temporal action localization ; Sports analytics ; Computer visionAbstract
This undergraduate thesis takes on the challenge of Player Centered Ball Action Spotting(PC-BAS) by the SoccerNet organization, where given a combination of tactical data, player tracking and broadcasts of games we are tasked with spotting players actions. A novel architecture is proposed which utilises a Heterogeneous Graph Transformer to combine the tactical data with the player tracklets and broadcast frames, and MS-TCN++ to temporally combine the graphs of different frames. In the training of the model we also used a context aware loss function(CALF) which was developed for another challenge of the same organization. This work enhances said loss function in order to account for the acting player. Furthermore, in order to provide explicit spatial awareness zonal nodes were introduced. To evaluate the performance of this Architecture we used the challenge's Codabench evaluation, where our full model achieves 39.9% micro-F1 and 25.2% macro-F1 on the VAL split, outperforming the graph-free ablation by
6.2 macro-F1 points, with the largest gains on relational action classes such as cross, block, and shot where inter-player context matters most.