Research And Application Of Algorithm For Motion Detection Of Racket Motion Video

With the development of artificial intelligence theory and technology,video-based behavior analysis technology has developed explosively,especially the innovation of machine learning algorithms and parallel computing technology,making deep video analysis possible.In the racket motion video,the technical movement has significant features,the racket motion background is relatively fixed,and the foreground characters in the racket motion video are also relatively simple,which has a prominent advantage in intelligent classification of the racket motion video.However,this type of motion video analysis also has many challenges,mainly manifested in how to effectively utilize the feature information of the image frame and how to accurately focus on the key information in the video image;In addition,there are also many problems such as a variety of video technologies,incomplete technologies,and difficulties in technology selection.Aiming at the above problems,this paper starts from the spatio-temporal correlation of video data,and introduces the attention mechanism based on CNN feature extraction to introduce the attention mechanism in semantic analysis and proposes a recurrent neural network based on attention mechanism.Each video frame is assigned a different weight to enhance the recognition effect in the captured video segment.Then the experimental performance analysis is carried out on the experimental data set.Through the experimental comparison method,the influence of the number of hidden units of LSTM,the value of dropout and the accuracy of different types of image input on the recognition accuracy of the shooting action are discussed.The experimental results of recurrent neural network based on attention mechanism show that the recognition effect of RGB image and optical flow image mixed input is better.Therefore,this paper proposes an improved dual-stream convolution hold motion recognition algorithm.On the task of capturing motion recognition,the batch normalization is combined with the Inception network as the basic network model,and the basic network model is used to construct the spatiotemporal dual stream convolution network architecture.At the same time,in order to make the model pay more attention to the racket action,the attention mechanism in semantic analysis is introduced into the recognition model,which helps the model extract more important information in the video data and pay attention to the key racket motion.Based on the relevant theories of video motion recognition in deep learning,this paper explores the temporal and spatial correlation of video actions,introduces the attention mechanism in deep learning,and analyzes the positive and negative hand movements in table tennis and tennis.The final practice shows that the improved Inception network can extract the feature information of RGB image and optical flow image more effectively.Introducing attention mechanisms allows the model to focus more effectively on key information in video images.The result of the fusion of the two-stream convolution network is input into the LSTM network including the attention mechanism,and the timing relationship between the images is fully utilized,so that the accuracy of the video classification of the action is improved.