Temporal Action Detection For Intelligent Broadcasting Of Sports Events

With the development of video technology,intelligent sports broadcasting has become an important trend in the field of sports broadcasting.Through artificial intelligence technologies such as deep learning,intelligent sports broadcasting automatically extracts information such as backgrounds,people,actions,behaviors,and relationships from multiple videos,and performs operations such as switching and replaying,thereby improving efficiency,reducing labor costs,increasing viewability,and promoting the development of sports events.However,currently there is a lack of available intelligent broadcasting technology in most sports events.In this article,we provide a detailed analysis of intelligent broadcasting technology in sports events,and establish a reasonable and reliable intelligent broadcasting system.The contents are as follows:(1)To address the low granularity of data set annotation in the sports field,this article establishes a high granularity multi-level table tennis data annotation standard,defining in detail the judging methods and differences for each category.Based on this,we establish a table tennis data set with high granularity features.Compared with general action positioning data sets,this data set has shorter action times,faster speeds,and smaller differences between actions,thus having higher granularity.(2)In order to solve the problem of slow model inference speed,this article adopts various improvement measures based on detailed analysis of the advantages and disadvantages of various networks,with the boundary-sensitive scheme as the basis.First,we introduce a multi-level boundary feature extraction layer to optimize the network structure and improve the inference speed.Second,we use global guidance loss to solve the problem of branch balance after the network becomes shallower.In addition,by using a method based on sparse convolution to extract confidence heads,the network can obtain a global view.Finally,we use a sampling optimization method based on difficult sample mining to further improve the performance of the model.The new model,Sparse Multilevel Boundary Generator(SMBG),compared with the newer method DBG based on the boundary-sensitive scheme,maintains the same accuracy on the standard data set ActivityNet1.3,but speeds up by 2-3 times,achieving good results and providing conditions for the application of intelligent broadcasting systems in sports events.(3)Finally,based on the high-granularity table tennis data set,we trained the optimized Sparse Multilevel Boundary Generator and achieved a testing accuracy of 47.5 AUC.Based on this model,we established an intelligent broadcasting system for table tennis events,which innovatively combines model recognition with prior knowledge of human broadcasting,has a concise and beautiful interactive interface,and can achieve automatic broadcasting function based on the input of multi-angle videos.