Research On Temporal Action Detection Based On Deep Learning

In past few years,along with the growth of network bandwidth,the progress of storage technology and the popularity of various video recording devices,people have been in an era of explosive growth of video data.Faced with massive video data,it is not realistic to rely on human resources to classify,identify and analyze video.Human is the dominant person of social life.Most of the valuable information in video is closely related to human behavior.The importance of temporal action detection based on artificial intelligence has become increasingly prominent.It has been widely used in many fields such as intelligent security,guardianship of patient and elderly,robot research and so on.How to analyze and process such a large amount of video data intelligently and mine the information becomes particularly important.Compared with the traditional machine learning algorithm,the biggest advantage of deep learning is that it provides an end-to-end training method for autonomous learning of data features,which need not be extracted manually.As a result,deep learning methods have made many breakthroughs in the field of artificial intelligence.Therefore,the application of deep learning technology in temporal action detection can greatly improve the performance of human behavior understanding in video,which has great research value.Based on the deep learning method,this paper designs a variety of section proposals network structure to obtain better temporal proposals.Multi-scale feature fusion strategy is adopted,context information and global information are added,and features are enhanced by deconvolution,which makes the features more representational,so as to make the network more accurate in the regression of temporal boundaries and classification of actions.In addition,the boundary trimming module is proposed as a general post-processing module to prune the temporal boundaries of the actions detected by the network accurately.The proposed scheme is simulated on the THUMOS 14-2014 data set,which achieves the highest accuracy compared with the research results in recent years,mAP@0.5 has reached 53.3%,and has a high recall rate and a low average number of proposals,the overall performance has been greatly improved.The universal boundary trimming module prunes the predicted time series positioning results of several existing excellent networks.The simulation results show that the proposed module improves the accuracy of the original network in varying degrees,and demonstrates the validity and versatility of the proposed module.