Implementation Of Distributed Detection System Of Illegal Actions In Video Based On Deep Neural Networks

In recent years,video surveillance has entered the city operation and people's lives with the acceleration of urban digital process.People and their action informations in surveillance video can provide data support for applications in multiple fields such as case detection,intelligent security and education supervision.The accuracy and efficiency of traditional manually analyzing surveillance videos relies on workers' personal abilities such as experience and stamina.Furthermore,limited workers cannot timely process the massive video data collected by monitoring equipments.Therefore,how to efficiently process the massive video data and accurately extract the information therein via computer vision is a current research hotspot.Multi-object action recognition is a senior task of computer vision that relies on multiple sub-tasks.How to combine these algorithms effectively and solve the problems of model compatibility,feature waste therein and how to satisfy the practical application requirements are the key points of this research.Aiming at the above problems and the demand of monitoring the use of mobile phones by employees in a pharmaceutical factory,this paper implemented a non-End2 End multi-object action recognition algorithm.Based on this,an End2 End multi-object action recognition algorithm that optimized the structures and merged the functions of algorithms and models was proposed.The contents are as follows:1.Non-End2 End multi-object action recognition algorithm.Most current researches on action recognition focus on single action in a video.Aiming at this limitation,this paper implemented a non-End2 End multi-object action recognition algorithm that combined and optimized YOLOv3 object detection algorithm,DeepSORT multi-object tracking algorithm and Pseudo-3D action recognition model on the input-output pipeline.This algorithm satisfied the efficiency requirements of quasi real time application based on certain computing hardware.2.End2 End multi-object action recognition algorithm.Aiming at the complex process and the computing and space resources waste of repetitive feature extraction operation in the non-End2 End multi-object action recognition algorithm.This paper used the ideas of feature sharing and 3D convolution to improve Faster R-CNN,proposed P3DRA(Pyramid 3D ROI Align)and TAN(Target Attention Network)toreplace the multi-object tracking algorithm,fused the object detection algorithm and the action recognition algorithm,and finally poroposed an End2 End multi-object action recognition algorithm.Experiments showed that,compared with the non-End2 End multi-object action recognition algorithm,the efficiency of this algorithm was greatly improved without affecting the actual application effect.3.System implementation and deployment.Considering the high concurrency of multi-channel video stream data inputs and the high throughput requirement of illegal actions detection function,this paper used Docker containerization engine,Kubernetes distributed container management framework,TensorServing model deployment server and QtGUI design platform to design,implement and deploy a stable,quasi real-time and scalable detection system of illegal actions in video.This system contains basic functions in the normal video monitoring system and can detect illegal actions of using mobile phones.