Events Detection Of Football Video Based On TensorFlow

As the largest sport in the world,soccer has attracted more and more attention.All kinds of soccer match have come into being,such as the European Championship,the World Cup,the Premier League and so on.However,soccer video has many characteristics,such as long duration and variety.Besides,people pay more attention to the highlights in these videos,such as shooting and corner in soccer match,which makes traditional video retrieval unable to meet people's needs.Therefore,how to quickly locate the video clips that users are interested in has become a research hotspot in current soccer video analysis.At present,although the method of soccer video event detection has achieved certain results,the process is complex,and the types of events that detected are limited.To solve this problem,we use deep learning method to construct event detection model to detect events contained in soccer video.The whole model is divided into two stages: the first stage is used to generate candidate event segments.Soccer videos to be detected are divided into specific length frame sequence,and use a sliding window to scan the video,the multiple frame sequences in a sliding window compose a segment.Each segment is a prediction unit and the sequences features in a segment are obtained through a three-dimensional convolution neural network,and then as the input of each time point of bi-directional recurrent neural network to further integration.Finally,output the event prediction of each video segment.The second stage is to further deal with the above results,remove all segments predicted as non-events,and set threshold to select the event segments which has higher probability according to the detection effect of various events,we can get the starting and ending position of the event by merging the selected segments,and do the classification and annotation,and finally output the complete event segments.In experiment,we use TensorFlow to implement event detection model and train and test it on soccer dataset.Experimental results show that the above method achieves certain results in soccer video event detection,and provides convenience for subsequent work such as video retrieval.How to extract effective features from different soccer events to improve the detection effect of individual events remains to be further studied.