Design And Implementation Of Staff Dressing Detection System Based On Image Segmentation

Nowadays,under the great supervision and promotion of national policies,enterprises pay more and more attention to the safety of Staff dress.Once security problems occur,they will bring huge property and economic losses to the company.The inspection and supervision of the staff’s dress is a problem that many enterprises need to solve.With the development of computer vision technology,many workplaces have now fully covered the surveillance camera,laying a good hardware foundation for the use of surveillance video data to achieve Intelligent inspection and supervision.This project is a cooperation project between our laboratory and the State Grid.It aims to reduce the labor cost of inspections,improve supervision efficiency,and effectively reduce safety by establishing a system that monitors and manages the safety of the staff in the entire workstation.The accident happened.This paper firstly analyzes several problems and difficulties in the implementation of the staff’s dressing detection system through monitoring video,based on the current technical conditions at home and abroad.and designs the system’s functional modules and the overall implementation process of the system for the actual scenario of the workstation application.This paper analyzes and compares the advantages and disadvantages of several target detection and instance segmentation techniques,and selects the Mask-RCNN algorithm which combines the best accuracy,speed and feasibility,It divides the input video frame picture and cuts out the human head,hands,upper body,lower body,and feet;then the image was pre-processed using technologies such as Batch Normalization.Finally,the VGG16 network was selected as the classification network,and the network structure was simplified in accordance with the actual application needs..Because the existing staff dress data set is very small,I organized and recorded more than 40 experimental videos of different scenes.After cropping and filtering,I obtained more than 100,000 valid samples of the head,hands,feet,upper body,and lower body as the data set.At the same time,it is combined with national grid video shooting on site for training and testing of convolutional neural networks.Through training,a model that can detect and classify the five types of safety wear of safety helmets,work clothes,work pants,insulating boots,and insulating gloves is obtained.The "dim" mode is designed for scenes with weak light to increase the system’s adaptability to different lights.After testing,one or more safety dresses can be detected and marked according to the user’s needs.And the system has adaptability to indoor and outdoor scenes and different shooting angles of the camera.For the scenes with poor light,the system’s "dim" mode has adaptability to scenes with weak light.Basically meet the needs of practical applications.