Research On Video Behavior Analysis Technology Based On Convolutional Neural Network

Video behavior analysis is an important research direction in the field of computer vision.The main research is to analyze and predict the behavior category of the target in the video by building a network.The network simulates the structure of human brain neurons through a computer.At present,artificial intelligence technology is increasingly mature in the field of image recognition.However,there are still deficiencies in the effective extraction of various behavior feature information in the video and the utilization of the time dimension relationship between video frames.This paper focuses on the relationship between the frames in the video in the time dimension,the spatial structure feature data information in the video and the effective extraction of the temporal motion feature data information.The specific implementation scheme is as follows:(1)Aiming at the complex diversity of video behavior feature information,in terms of feature extraction of behavioral information in videos,a video behavior recognition algorithm based on spatio-temporal dual-stream heterogeneous convolutional neural network is proposed.The algorithm is improved on the basic structure of the traditional dual-stream network,and it extracts effective spatial information and motion information from the behavior information of the video in two channels.When extracting spatial information,considering the complexity and diversity of spatial information in RGB video images,the Dense Net structure is used between the network layers to construct a dense network to extract spatial information.When extracting motion information,considering that the feature information contained in the optical flow diagram is relatively small,BNInception network is used to extract motion information.The experimental results of this paper on the UCF101 video behavior analysis database show that the accuracy of the video behavior recognition algorithm based on this network structure reaches 85.5%.Its feature extraction capability surpass over the video behavior recognition algorithm based on traditional dual-stream network.(2)Considering the correlation between video frames in most videos and the ability of the network to effectively extract feature information in video behavior analysis,this paper optimizes the video behavior analysis algorithm based on spacetime dual-stream heterogeneous convolutional neural networks.A video behavior analysis algorithm based on time-correlated sampling dual-stream heterogeneous grafting network is proposed in this paper.In the video sampling stage,this paper adopts the temporal correlation sampling of video frames based on the TRN algorithm.Using this method to sample the video can make full use of the relationship between the video frames.Based on their changes in the time dimension,many easily confused behaviors can be well distinguished.In the feature extraction stage,this paper improves BNInception and Dense Net based on Filter Grafting.In order to improve the efficiency of the network convolution kernel,this technology grafts external information or network internal information to the location of the invalid filter.After reactivating them,it can effectively improve the network model's ability to express feature information.The experimental results of this paper on the UCF101 video behavior analysis database show that the video behavior analysis algorithm based on the time-correlated sampling dual-stream heterogeneous grafting network proposed in this paper has a recognition accuracy of 89.3%.This algorithm further effectively improves the accuracy of the algorithm.