Tennis Video Summarization System Using Recurrent Neural Network

Video has quickly become one of the most common visual information resources,and sports event video is an important part of the video field.For those athletes who want to become stronger,it is an indispensable means to watch videos to improve their own leaks.However,unnecessary clips in the video when watching the video will waste a lot of time for the athletes,which is very unfavorable for the daily training of the athletes.Therefore,it is very efficient and convenient to develop a smart tool that can analyze and understand video content and cut out unnecessary video clips in the video.Video summary refers to extracting representative key frames from the original video,reducing the original video to a shorter video while saving the most important and representative information in the video.A video summarization is a problem of sequence structure prediction.Its input is a series of video frames,and the output is a binary value used to indicate whether a video frame is selected.Since the hardware computing power has been greatly improved,the neural network algorithm has gradually become an important method to solve related problems,especially the long short-term memory network in the recurrent neural network,which can synthesize all the data of the original video to determine whether the current frame is selected,thus becoming the most central method for dealing with sequence structure prediction problems.This paper aims to design a video summary system for tennis based on the recurrent neural network,using long short-term memory networks to model the variable range dependencies of long-term and short-term interleaving.By combining with the pedestrian detection algorithm based on Faster R-CNN,the useless information in the video frame is filtered out.At the same time,combined with the Flownet2.0-based optical flow estimation algorithm,the original color image is converted into an optical flow image.The accuracy of the tennis video summarization algorithm is improved by combining the information of the color image and the information of the optical flow image.At this stage,our method is trained on a 2153 minute dataset,and the accuracy rate on the 461 minute testset reaches 81.5%.