Research On Deep Learning Based Automatic Description Method For Free Exercise Video

Under the background that modern people increasingly pay attention to health and respect sports,the amount of data and audience of sports videos is growing rapidly,and its potential application value has attracted widespread attention from scientific researchers and industry.The research content of this article is automatic understanding of floor exercise video.The so-called floor gymnastics video comprehension refers to observing the athlete's complete set of movements in the video to generate a professional term for the set of movements performed by the athlete.It has extensive application value in sports analysis,automatic interpretation,and sports instruction.The research on the automatic understanding of the content of floor exercise video in this article is more specifically the automatic understanding of the human movement in the exercise video.This article will combine computer video and deep learning-related knowledge to implement intelligent tagging and representation of specific human motions present in video sequences.The main work of this article is as follows:Automatic description of floor exercise videos based on long-term and short-term memory networks.In the classic video description model S2 VT,a long-short-term memory network is used to learn the mapping relationship between word sequences and video frame sequences.This article introduces the attention mechanism for improvement,highlighting the importance of determining the key frames of floor exercise.In this paper,a data set of floor exercise decomposition for professional events is established.Experiments are performed on MSVD data and self-built data sets,and the planned sampling method is used to eliminate the differences between training and predictive decoders.Experimental results show that the improved method in this paper can improve the accuracy of automatic description of floor exercise video.At the same time,in this experiment,the effects of different convolutional networks on feature extraction are also compared,and the effect of feature extraction networks on the automatic description of floor exercises is analyzed.Automatic description method of floor exercise video based on 3D convolutional network and multi-label classification.A set of floor exercises consists of multiple disassembled moves.In the work of this paper,a classifier with a single decomposition action is constructed to convert the automatic labeling problem of floor exercises into a multi-label classification problem.As can be seen from the previous chapters,as the depth of the feature extraction network increases,the experimental effect increases,but the 2D convolutional neural network will lose time information when extracting features.Feature extraction.Perform multiple binary classifications on the extracted features to achieve the goal of multiple classifications.In order to form a comparative experiment,the results of classification are randomly combined into a sentence and compared with the results of automatic description to verify the effectiveness of the method.