Research On The Surveillance Video Image Translation And Generation Algorithm Based On Deep Learning

Supervision places are an important part of my country’s judicial system and one of the important state apparatus.Due to its particularity,supervision places have higher requirements for the construction of safety and protection systems than other industries,In recent years,with the rapid development of artificial intelligence technology,more and more intelligent technologies have been used in the security system of supervision places,such as automatic target classification,identification of illegal areas,and identification of fights.Combining the "technical prevention" system on the basis of the traditional "manual prevention" system in supervision sites can significantly improve the quality and efficiency of supervision.It has become a trend to apply deep learning,cloud computing,big data and many other emerging technologies to the construction of security systems in supervision places.However,the construction of an intelligent security protection system in a supervision places requires a large amount of data.In view of the particularity and confidentiality of the supervision places,it is obviously unrealistic to obtain a large amount of relevant data.According to the requirements of the national key research and development project "Research and demonstration of key technologies for intelligent monitoring,early warning and prevention in supervision places",this thesis studies the problems of lack of scene data and single scenes in supervision places to improve the functions of the intelligent security system in supervision places.Video surveillance is the foundation of the security system in supervision places.Traditional video surveillance relies on manual surveillance,and there are too many surveillance pictures in the supervision places,which makes surveillance inefficient.The intelligent video surveillance system can make the surveillance mode change from passive to active.However,there are very few relevant data on supervision places,which makes it difficult to research and improve the security system of supervision places.In order to solve the above problems,this paper studies how to generate images of supervision places based on a very small amount of available real data,design algorithms to improve the quality of generated data in specific scenarios,and generate abnormal behavior scenarios of detainees based on this algorithm,and finally design an abnormal behavior detection system and use the generated data to test its function.Taking the project requirements as the background,this thesis solves the problem of missing data in supervision places,such as seasonal transfer.This thesis is based on the dataset with universal characteristics of the target season,uses semantic segmentation algorithm to perform semantic information statistics on the dataset,builds an image semantic dictionary,and proposes an image matching algorithm to improve the important components of the image-to-image translation algorithm.In order to improve the effect of seasonal transfer of supervision places,this thesis proposes a static background translation algorithm based on image semantic adaptation.After analysis,the effect of the image-to-image translation algorithm is positively correlated with the degree of semantic matching between the initial image and the learning target image.For the static background seasonal transfer task of the supervision places,the semantic matching algorithm is used to select the best target for learning in the constructed semantic dictionary,which can improve the quality of seasonal migration.Not only that,in addition to the existing qualitative and quantitative evaluation criteria,this thesis designs a color feature-based background seasonal transfer quality evaluation criteria to evaluate the generated images in terms of color features.After experimental verification,the effect of the algorithm proposed in this thesis is significantly improved regardless of the existing criteria or the criteria based on color characteristics.In order to solve the problem of character uniform errors in the seasonal transfer of the static background of the supervision places,this paper proposes a translation algorithm for the supervision place based on the separation of dynamic and static targets.On the basis of the semantic-adapted background translation algorithm,the dynamic target and the background in the view of the supervision places are independently learned on different branches.After the learning,the branches are merged according to the coordinates.Not only that,in order to evaluate the seasonal transfer effect of the view of the supervision places,in addition to the existing qualitative and quantitative criteria,this paper designs a target detection-based quality evaluation criteria for the generated images.After experimental verification,whether under the existing criteria or the criteria based on the target detection,the algorithm of the supervision places based on the dynamic and static target separation can solve the problem of the translation error of the character uniform.Finally,considering the actual needs of the project,in order to solve the problem of the lack of scenes in the supervision places,taking the abnormal behavior of the detainees as an example,this thesis designs an abnormal behavior detection system in the supervision places based on the sample-assisted generation.Video sequence samples were generated for the three typical abnormal behaviors of detainees:entering the illegal area,deregulated behavior,and fighting.Then,each module of the system was tested separately to meet the project requirements.