| In recent years,the country has continuously strengthened its construction in the field of security.It installed a large number of monitoring devices,causing monitoring video accumulation.However,traditional visual monitoring methods cannot effectively use massive monitoring videos.With the rapid development of artificial intelligence,deep learning algorithms can be used to analyze video content in real time.It can detect abnormal information,so as to carry out risk prediction and construct an intelligent monitoring system.In order to make effective use of the security monitoring video,this paper presents a study on the automatic detection of character identity attributes in video using the deep learning algorithm.The main research contents and conclusions of this article are as follows:1.Aiming at the need of judging the identity of the person under the current monitoring environment,a method based on the automatic recognition of the character identity of the deep learning was proposed.Because the traditional target detection algorithm needs manual design to extract image features and the video detection can not reach the real-time problem,this paper designs a recognition method based on deep learning of Faster RCNN.The paper first designs the flow chart of character identity attribute based on Faster RCNN algorithm,and studies the overall framework of Faster RCNN algorithm in detail.The framework is composed of RPN candidate box extraction module and Fast RCNN detection module,and introduces the network structure of the two parts of the module.Then the training parameters are set according to the research objectives to complete the regional proposed network design.Finally,the paper elaborates the training method and steps of character identity detection based on Faster RCNN algorithm2.Build a deep learning platform and build a data set for essay research.Through the comparison of various deep learning frameworks,the Caffe framework is selected as the training platform.The paper uses the prison surveillance video as background to collect surveillance video for thesis experiment.Condensing video through video enrichment techniques to select videos of people’s activities.The labeling of two person identity attributes in the data set is accomplished by image annotation.3.Due to the small number of data set classification categories,the ZF network is selected as the initial network structure of the regional recommendation network.By setting the number of iterations and the learning rate for network leaming,the final classification network model is obtained.Finally,the paper compares the experimental results with the results of training on the Fast RCNN network.In the classification accuracy rate,the accuracy rate of the two categories was increased from 0.670 to 0.828,from 0.579 to 0.708,and the detection speed was also increased from 2s to 0.04s.The test results show that the trained model successfully identifies both the prisoner and the police in the monitoring and achieves the purpose of real-time testing,which verifies the feasibility of the algorithm. |
