Application Research Of Broad Learning System In Dock Security

Driven by the development of related fields such as machine learning,computer vision,image processing,and pattern recognition,video image recognition technology has been gradually applied to many walks of life in society.Among them,the field of shipping logistics also uses video surveillance as the main management tool for its safe production.Many studies have proposed to implant video image recognition algorithms into video surveillance systems to achieve intelligent supervision of Container Freight Stations(CFS).However,the container freight station has a busy logistics operation and various operation equipments.For supervision,in complex environments with relatively few safety hazards,how to realize the hazard source identification on the job site is a challenging problem.As far as the personal safety of workers is concerned,wearing safety helmet is the most direct and effective method.Therefore,supervision safety helmet wearing has become the most important part in safety production.In order to solve the problem of resource waste and low monitoring efficiency caused by the manual guarding method in the safety inspection of the dock security,this thesis focuses on the CFS,an important terminal operation site,and conducts in-depth research on the identification algorithm module in the CFS operator’s helmet detection system.Firstly,the image is extracted from the video.After the image is grayed out and morphologically preprocessed,the Visual Background Extractor(ViBe)and Gaussian mixture model(GMM)are used to extract the moving target from video image,which is used as a candidate set for image detection.Experiments show that the ViBe algorithm is greatly affected by "ghost" phenomenon in the actual extraction process.Therefore,this study selects GMM algorithm to detect the motion foreground of video images,then combines some images positioning technology to perform precise positioning and segmentation of the image,and divides the extracted image into three part,namely a CFS background,an operator wearing a helmet,and an operator without a helmet,to construct an image database.Then,Convolutional Neural Network(CNN),Support Vector Machine(SVM)and Broad Learning System(BLS)are used to conduct comparative experiments in the "Multistage" and"End-to-end" identification modes.The experiment shows that the helmet image recognition algorithm is generally more stable in the "End-to-end" mode,with higher recognition accuracy and shorter model training time.BLS in the“End-to-end" mode can achieve the highest accuracy and the shortest model training time in the helmet detection task,which lays the foundation for the subsequent application of the helmet wearing state recognition technology to the CFS surveillance system.Finally,in the task of detecting whether the operator wears a helmet in the CFS,BLS in the "End-to-end" detection mode is selected as the image recognition classifier for the helmet detection.The image of the operator who does not wear a helmet in the CFS is marked accurately and quickly,which makes it convenient to intelligently identify the helmet image under limited hardware conditions and the effect proves to be good.