Research On Key Technologies Of Video Monitoring In Workshop Production Safety

Production safety is a necessary condition for the survival and development of enterprises.In large workshops,the production safety of employees should be given priority.At present,the real-time tracking and positioning of workshop employees and whether the wearing of helmets and workwear conforms to the regulations mainly depend on manpower completion,which not only wastes a lot of labor costs but also is inefficient.With the development of deep learning in the field of intelligent video monitoring,intelligent automatic analysis and recognition of video images can be achieved to improve work efficiency.Through intelligent video monitoring technology to replace manual detection,can truly achieve safety production automation,high efficiency management.For this purpose,this paper proposes the research on key technologies of workshop production safety video monitoring based on deep learning.The work and achievements of this paper are divided into the following five aspects:(1)Firstly,this paper designs an improved YOLO v5 s model to solve the problem of inaccurate prediction box and low detection accuracy caused by the target detection algorithm under large area occlusion of pedestrians.In order to enhance the ability of pedestrian feature extraction,this paper introduced ECA-Net attention mechanism into the neck network.Then,considering the influence of model size on real-time performance,this paper introduces an atrous spatial pyramid pooling module into YOLO v5 s backbone network to reduce the loss of context information and the computational load of the model.Finally,combined with CIo U,a new non-maximum suppression method is proposed.Experimental data shows that the proposed model can effectively improve the accuracy of pedestrian prediction box in the case of large area occlusion.(2)In terms of pedestrian tracking,considering the model size and real-time tracking,this paper designed to replace the original appearance feature extraction network with Shuffle Net V2 network and combined with the improved YOLO v5 s detection algorithm to form a lightweight Deep SORT pedestrian tracking model.Experimental data on large the tracking data set shows that,the proposed model has improved tracking speed and accuracy.(3)In terms of safety wear detection,this paper proposes an object detection method to improve feature fusion.In this method,the simplified Bi FPN is used as the neck feature fusion structure of YOLO v5 m,and the CBAM module is added to the Bi FPN behind the output layer and in front of the YOLO v5 m prediction layer to enhance the feature capture and expression ability of the model.In the model training stage,a novel α-EIo U loss function is used in this paper.Finally,the experimental results on the self-built dataset show that the proposed method can effectively improve the detection accuracy of wearable categories.(4)Aiming at the requirements of workshop intelligent management and workshop target pixel-level segmentation,a U-Net semantic segmentation model based on RepVGG lightweight backbone network is designed.Then,pyramid squeeze attention mechanism is added in the U-net fusion stage,and the combined loss function is used in model training to achieve lightweight and fine segmentation effect on the self-built small sample workshop dataset.(5)Finally,the workshop production safety video monitoring system is designed and implemented with Py QT5 development tool.The functional modules of the system mainly include login and registration,pedestrian tracking and detection,safety wear detection,and semantic segmentation with the workshop scene.Through experimental tests,the system has the characteristics of light weight,easy to install,occupy small space and so on.At the same time,the interface visualization degree is high,can meet the needs of human-computer interaction.