Design And Development Of Intelligent Video Inspection Assistant System Based On Improved YOLO

Our country is a giant industrial production country.In chemical companies,regularly inspection of the behavior of the equipment and factory staff is an indispensable part among the enterprise operation.Nowadays,most inspections in the chemical industry still rely on manual labor.In addition,it is easy to "cut corners and materials" during the inspection process,and the problems of improper inspection and insufficient time also occur from time to time.Accidents are extremely prone to occur in the long run.Therefore,intelligent inspection system is urgently needed to assist or replace manual inspection.In recent years,artificial intelligence technology has developed rapidly,and the concept of smart plants came into being.Scholars have gradually been deeply in-depth research on industrial intelligent inspection.In the context of the daily inspection of chemical enterprises,this paper fuses the target detection technology.By continuous improvement,the YOLO algorithm is optimized,and the accuracy is improved.Finally,the improvement algorithm is embedded in the intelligent video inspection assistant system to meet the requirements of the auxiliary manual inspection.The main research contents of this article are as follows:(1)Analyze the research status of intelligent monitoring and unsafe behavior detection at home and abroad,combining the practice experience in chemical companies to find out the pain points in daily inspection,and conceive solutions to this,and initially design the system structure.(2)By consulting a large number of documents,master the working principles of popular target detection algorithms and analyze the advantages and disadvantages of various algorithms.Considering that the inspection process needs to meet the real-time performance and the recognition accuracy rate,this paper finally chooses YOLO as the prototype network for improvement and optimization.(3)The algorithm has been improved,and finally obtained the DDCN-YOLO model.First,from the training data integration,use the data enhancement and the MOSAIC method to enrich the background of the data set to strengthen the target resolution of the network in the complex environment;secondly,the DBSCAN-KMeans ++ cluster algorithm uses Anchor Box in the model.Re-clustering to increase the identification of small targets in the image;again,in order to prevent image characteristics from gradually weakened with the network depth,this article introduces DENSENET in the network;in addition,this article will change the variable convolution DCN Combined with the original network to increase the perception of the field;Through experiments,the accuracy is 2.33% higher than YOLOV4,and the leak detection rate of small targets is reduced.At the same time,the model performance is verified by the longitudinal and lateral contrasts.(4)The system is developed using the Python language and conducts model training under the Tensorflow 2.0 deep learning development framework,with training data sets from network climbing and manual shooting.After training,the identification of four insecure behaviors of personnel intrusion,flame,smoking and unwrapped hard hat.The production of the system human interactive interface is used in the PYQT5 development framework.The system background uses MySQL relational database management information,and the system is developed by PyCharm IDE.