Action Recognition Based On Deep Learning

In recent years,with the continuous development of mobile Internet and mobile smart terminals,video data has experienced a spurt of growth.Using computers to process these data efficiently has extremely high application value.Human action recognition has been a research hotspot in the field of computer vision in recent years.The rapid development of deep learning has greatly promoted the research in computer vision related fields.Although deep learning has achieved fruitful results in single image recognition,further research and development are still needed in the recognition of video streaming media content.The specific performance is that the recognition accuracy is not high,the recognition granularity is too coarse,and the recognition speed is slow.Based on deep learning technology,this paper provides algorithm schemes and engineering implementations for shooting motion detection tasks.The main content and innovations of this article are summarized as follows:Aiming at the problem of locating target actions from long videos,this thesis gives specific implementations from two algorithm frameworks.The first scheme uses the RPN layer to estimate the boundary of the shooting action through regression,and judges the confidence that the action within the boundary is the shooting action through the method of classification;the second scheme uses the sliding window method,which uses a fixed-size window in the long video Sampling in the middle,and then use the classification network to judge each window to finally judge the position of the shooting action.Aiming at the problem of inaccurate detection of shooting motion boundaries,a structured detection branch is added to the original network.The content detection subnet focuses on intra-frame motion detection,and the structured detection subnet focuses on the initial posture detection of the shooting motion.Finally,the detection results of the two branches are combined according to a certain proportion to determine the specific location of the shooting motion.In summary,by studying the key technologies of action localization,this paper realizes the basic function of video action extraction for the specific scene of shooting motion.At the same time,for the situation that the recognition granularity is coarse and the action localization is poorly structured,Structured detection network module is proposed.This module greatly improves the ability to recognize fine-grained actions,and finally realizes the algorithm as an engineering service,so as to provide effective solutions for the research and practice of gesture recognition algorithms in specific scenarios.