| With the development of Internet,network transmission and information technology,video resources are more and more abundant,network traffic gradually tends to video,and network video is more diversified.Today,with the explosive growth of network video traffic,video content security problems gradually emerge,and have been highly concerned by the national government and all sectors of society.Violent video,as an important part of it,can not be ignored.How to detect violent scenes is a challenging task.At present,the technology of violence video detection based on deep learning is mainly divided into the following methods:(1)after the video is divided into frames,3D convolutional neural networks(C3d)network is used for classification;(2)convolutional neural network(CNN)is used to extract the features of video frames,and then long short term memory(lttmm)network is used,LSTM)is used to model the time sequence signal,and then classify it by classifier.(3)the dense trajectory features and CNN extracted features are fused to classify.(4)the features of visual channel and audio channel are fused and classified.There are two common problems in the visual channel of the above methods:(1)the trained model is sensitive to the background information and ignores part of the foreground information.The human brain’s judgment of violence mainly depends on the foreground information,and the judgment decision of the model is not consistent with the human brain’s cognition,which will lead to the poor generalization ability of the model(2)It is difficult to model in the crowded scene,and some violence video detection algorithms can not locate the specific human area to judge the violence.To solve the above problems,we first use the fast r-cnn + C3 d model for mask experiment to verify the existence of the problem(1).Further,we use the gradient weighted class activation mapping(Grad CAM)to visualize the impact of some regions in the video frame on C3 d violence detection.According to question(1),we first use fast r-cnn for human recognition,and then erase the background information outside the target person to train C3 d network.We find that the preprocessing of background information can improve the generalization ability of the model to a certain extent.Furthermore,we propose F3 D model to improve the accuracy of violence detection.Then,for problem(2),we use the improved yolov3(you only look once V3)model to test on the violentcrowd data set.The experimental results show that it is difficult to detect the target person with the traditional target recognition method in the dense crowd scene,and there is a problem that the specific human region cannot be located.Therefore,we use csrnet(network for condensed scene recognition)to generate a crowd density map to represent human body region information,and fuse it with the information extracted by reconet(a low rank to high rank context reconstruction framework)to identify crowd violence in dense scenes.Finally,we apply the proposed method to three binary violence video detection datasets to verify its effectiveness.For crowd violence detection in sparse scenes,we use hockey fits data set and violencemovies data set.To solve the problem(1)we apply the proposed human region information fusion model F3 D to the above two datasets for 50% cross validation experiments.The experimental results show that the average accuracy is 96.25% on the hockey fits dataset,which is 0.54%higher than the baseline model,and 98.97% on the violencemovies dataset,Compared with the baseline model,it was improved by 0.2%.For crowd violence detection in dense scenes,we use the violet crowd dataset.To solve the problem(2)we apply the feature fusion model recomap3 d based on density map and reconet attention mechanism proposed in this paper to the violentcrowd dataset for 50%cross validation experiment,and the average accuracy is 99.33%.Compared with the baseline model,the accuracy is improved by 0.93%,and the highest known accuracy is obtained. |
