| The characteristics of video data,such as large amount of data,unstructured and diverse content,make the classified retrieval and effective management of video a long-term problem.With the rapid iteration of mobile communication technology and the development and popularization of intelligent devices,the video data in modern society has increased sharply,and now it has occupied most of the traffic share of the Internet.Compared with words and pictures,video can convey information and express emotion more accurately.Therefore,people are more and more willing to use all kinds of short video applications to create and share their own videos.More and more huge video data,so there is an urgent need for effective video understanding and video analysis technology to process.Motion recognition is the core task in the field of video understanding.The research in this direction can not only explore a better video understanding model,but also promote the development of other tasks.However,most of the traditional video action recognition frameworks use the original RGB video,optical flow and three-dimensional convolutional neural network for recognition,which makes the action recognition task encounter great storage and performance obstacles in the mobile terminal,and it is difficult to be suitable for the application scenario of the mobile terminal.To solve the above problems,this paper proposes an action recognition algorithm based on compressed video.Using the key frames,motion vectors and residuals in compressed video,the model with greatly reduced parameters is trained by two-dimensional convolutional neural network.Finally,action recognition is carried out,so as to greatly reduce the storage cost and computing cost.In addition,this paper also introduces the preprocessing of motion vector,attention mechanism and feature fusion mechanism based on recurrent neural network structure to optimize the insufficient timing information fusion of the algorithm,improve the recognition accuracy and enhance the robustness of the algorithm.Based on the improved compressed video action recognition algorithm,this paper designs a set of action recognition system running on Android mobile terminal.The function and performance test of each module and page of the system shows that the system can effectively realize the action recognition of local compressed short video at the mobile end,and label and classify the short video according to the recognition results,which is fully effective and feasible. |
