Research On Video Acticon Recognition Based On Video Content

With the proliferation of video information and the rapid development of machine vision,people have higher and higher technical requirements in the field of video processing.Video action recognition is a branch of video understanding direction that has very important research significance.In this paper,the video action recognition algorithm is studied in detail.It aims to form a more discriminative video action recognition system by analyzing the spatio-temporal information of video to form more effective video features.Firstly,a video action recognition algorithm based on spatio-temporal convolutional network is designed,which deals with video action recognition as separate acquisition of video spatial and temporal information.Based on the two-stream convolutional network,the effects of optical flow image and saliency image as the network input on the action recognition effect are compared and analyzed.At the same time,through the analysis of video spatio-temporal information and 3D convolution,a spatio-temporal convolution network with dual branch structure is constructed,which can obtain the spatial and temporal information of video in an independent and unambiguous way,and can effectively improve the results of action recognition.Secondly,a video action recognition algorithm based on recurrent neural network for feature fusion and video segmentation is designed.The video action recognition work is processed according to time series information.Based on the recurrent neural network,the action recognition algorithm combined with convolutional neural network is firstly implemented,and the network type and sampling length are compared and analyzed.Furthermore,based on the neighboring feature averaging and 3D convolutional networks,a recognition algorithm based on feature fusion is proposed.Finally,a action recognition algorithm based on video segmentation is proposed,and the effects of fusion function and video segment number on action recognition are compared and analyzed.Finally,a video action recognition algorithm based on 3D convolution dense network is designed to process video action recognition to simultaneously acquire video spatio-temporal information.Based on DenseNet's dense hierarchical connection method and 3D convolution operation,a 3D convolution dense network is constructed.At the same time,by studying the scale of 3D convolution processing video sequence,a multi-scale time input conversion layer is proposed,which can be densely constructed.And a multi-scale three-dimensional convolution dense network that efficiently acquires spatio-temporal information of video,and can effectively improve the action recognition.