Research On Highlights Detection In Football Videos Based On Rules And Transformer

Football video highlight event detection is of great significance for video retrieval,video summary generation and team tactical analysis.The existing football video highlight event detection methods have made some progress,but there are still some problems:artificial rule-based methods need to manually define semantic rules,but due to the low universality of rules,event detection is often inaccurate;and Methods based on convolutional neural networks require pre-training on a large number of training sets,which needs a large amount of calculation.Aiming at the above problems,through in-depth analysis of the characteristics of football highlights,a football highlight event detection method combining football rules and Transformer is proposed.Through research and analysis,the highlights in football videos are defined,and the rules of highlights are summarized and verified.The football video is preprocessed according to these rules,and the prediction unit is obtained for further processing.Aiming at the problem that Vision Transformer cannot be directly applied to video tasks due to its high computational cost,a video classification model GLTrans that only relies on attention mechanism,namely global-local Transformer model,is proposed to further classify and filter prediction units.The model divides each frame of image into several image blocks equally,and can better model the temporal and spatial information of events through the global and local attention mechanism.Use global and local attention models to further filter highlights and refine temporal boundaries for more accurate detection results.The experiment uses the SSET football dataset to train and test the model.The results show that the proposed football video highlight event detection method has excellent performance.Compared with Time Sformer,the F1-measure is improved by 7.09%,and the time series boundary is closer to the real event.However,the model only uses the characteristics of the event itself,and how to combine the contextual semantic information of the wonderful event needs further research.