Prison Video Surveillance Sorting Research Based On Deep Learning

The continuous development of artificial intelligence technology provides a lot of help for the construction of intelligent prison,and provides a lot of convenience for the prison supervision department which requires high real-time monitoring.Compared to the traditional video monitoring recommendations,based on deep learning of intelligent recommendation technology to reduce the intensity of administrative personnel to monitor video to watch at the same time for staff to watch more appropriate surveillance video,which to a certain extent,reduce the workload of prison management personnel,improve the efficiency of the prison to personnel activity regulation.At present,the recommendation system technology is mainly applied in the field of e-commerce,and almost all the technical solutions are personalized recommendation solutions based on the matching of the user's identity and the characteristic information of the recommended object.However,in the application scenarios of prison video surveillance recommendation,the operator information is usually uncertain,and the identity characteristics of the user and the object characteristics of the operation object cannot be obtained,which makes it difficult to apply collaborative filtering methods in the traditional recommendation technology to the recommendation of surveillance video.How to make better use of the sequence of monitoring operation objects is particularly important.In recent years,many researchers use the hidden state of recurrent neural network to save the important historical information of sequence,and the recommendation scheme which focuses on the different importance degree of sequence items with the attention mechanism has achieved obvious results in the problem of sequence recommendation.However,the existing recurrent neural network methods based on attention mechanism do not make explicit use of the context local characteristics such as the mode of object transfer to the click sequence,the interval length of the click operation object at different moments and the different time distances.In order to solve these problems and combine with the actual situation in the prison surveillance scene,this thesis makes further improvement on the basis of the existing results,and the main innovation points are as follows:Firstly,we embed the temporal transfer mode of operation objects in the operation organization,model the transfer mode and dependency relationship of operation objects by introducing the method of graph structure data embedding with position coding.Meanwhile we use the threshold function to control the retention and abandonment of local features.This method can consider the deeper temporal dependency and transfer information between different operation objects,so as to more accurately represent the embedded representation information of operation objects and improve the final recommendation effect.The experimental results on real prison datasets also show the effectiveness of the embedded representation method.Secondly,the traditional methods only consider the temporal information and context information,ignoring the different influence of different time distance features on the final result.This thesis also proposes a long-short term attention network to explore the influence of user's local behavior on the global characteristics.In real tasks,user's intention is time shifting.We adopt different attention mechanisms to obtain user's intention in different time range.The segmented attention network in this thesis combines the advantages of recurrent neural network and attention network.It considers the different dependence of long-term and short-term operation objects on the global organization at the same time,and express the influence of different time interval object groups on the operation organization in different characteristics.The experimental results on two prison datasets also show the effectiveness of the network structure.