Research And System Implementation Of Behavior Recognition Method Oriented For Computer Room Security

In recent years,the 2D convolutional neural network as a typical machine vision processing algorithm model has achieved great success in image target recognition applications with its powerful feature learning and transfer learning capabilities.However,this type of algorithm model is only suitable for target recognition in static images,and cannot meet the needs of complex time-series behavior recognition in applications such as security monitoring and behavioral regulation supervision of specific types of work.Through cascading 2D convolutional neural networks and recurrent neural networks or using 3D convolutional neural networks can make models obtain the ability to recognize time-series behaviors,but these networks are large in scale,complex in training process,and difficult to connect to practical applications.With compliance of management requirements of the State Grid Center’s computer room,a deep research of complex time sequence behavior recognition method and a full design of efficient and system combining behavioral graphs and 2D convolutional neural networks are successfully illustrated in this thesis.A set of effective solution is also provided for the application requirements of complex time sequence behavior recognition in real scenarios.In this thesis,deeply investigations of the complex time-series behavior recognition requirements,which commonly exist in computer room security supervision are conducted,as well as detailed analysis of the performance of behavior recognition methods based on deep learning for application difficulties.At the same time,a general behavior recognition method assisted by behavior graphs and a set of informatization Computer room security supervision system scheme are also provided in this thesis.Based on the whole analysis above,a lightweight behavioral element collection algorithm architecture that can be deployed at the edge is designed,which can also accurately identify complex timing behavioral elements.Based on the edge computing architecture,combined with the server-side and client-side management platforms,this thesis is successfully completed with the design of the information-based computer room security supervision system.The innovations of the thesis include: 1)In view of the lack of efficient algorithm models for complex time series behaviors,a generalized and efficient identification method combining behavioral graphs and 2D convolutional neural networks is proposed;2)Cloud computing solutions require large bandwidth,poor timeliness,and operational Due to the problem of high cost,an edge computing solution based on a lightweight behavior element collection algorithm architecture is designed to realize real-time edge recognition of complex timing behaviors.After testing,the computer room security supervision system designed in the thesis runs stably,and accurately identifies employee behaviors such as double-opening safes.In the simulated computer room environment,the accuracy of identifying the violations in the computer room has reached 93.5%,which better meets the real-time requirements of the computer room monitoring system.The system power consumption also meets the edge operation requirements,and it has good human-computer interaction functions.The performance indicators of the system have reached the expected research goals.