Research On The Detection Model Of Healthcare Workers’ Protective Equipment Donning Normality Based On Deep Learning

The irregular use of personal protective equipment(PPE)seriously affects the occupational health and safety of healthcare workers,especially in major public health emergencies such as COVID-19.However,the existing manual monitoring methods suffer from high costs,low efficiency,and many loopholes,which cannot meet realistic needs.Recent advances in computer vision provide an automated monitoring platform for occupational safety and health monitoring of workers in high-risk areas.Although a large number of studies have demonstrated its application potential,the lack of datasets leads to few systematic studies on the detection of normative PPE donning by healthcare workers.To address the above issues,this thesis investigates intelligent detection models based on deep learning for the normative donning of healthcare workers’ protective equipment.The details include:(1)Medical protective equipment donning normative standards and dedicated dataset construction.To address the problem that there is no public image dataset suitable for studying the normative donning PPE of healthcare workers,we face the training needs of visual detection models,and based on the developed medical PPE donning normative standard,by cooperating with professional healthcare workers,collecting data in a combination of field acquisition and laboratory simulation,performing bounding boxes and points annotation,making offline and online data enhancement to construct a medical staff dress code dataset MSDCD in a structured scenario.The dataset contains a wide range of PPE misuse and normative use samples that may occur in medical operation scenarios and contains a total of 6,000 pre-processed images.(2)A PPE-donning normative detection model for healthcare workers based on YOLOv4.In response to public health emergencies,the PPE donning detection task for healthcare workers has a dual demand for model performance and efficiency,especially monitoring efficiency and real-time,we propose a PPE donning detection model MSPPEYOLOv4 based on the One-stage object detection algorithm.This model simultaneously locates and classifies six classes of PPE objects,including medical protective hats,glasses,masks,clothing,gloves,and shoes,and recognizes the normality of donning multiple pieces of equipment end-to-end.The results of the analysis on the self-built dataset MSDCD show that the proposed MSPPE-YOLOv4 model achieves a relatively better balance of higher accuracy(84.14%)and less processing time(42.02ms)than the typical representative algorithms of Two-stage and One-stage.(3)A PPE-donning normative detection model for healthcare workers fusing human keypoints.For complex scenarios such as protective glasses held in the hands of workers and not properly worn,a single object detection method may result in wrong or missed detection problems,we further propose a deep learning detection model that combines human keypoints and object detection.This model uses YOLOv4 as the baseline model for PPE detection and MobileNetv3 as the backbone network to reduce the computation to achieve model lightness;HRNet is the benchmark for keypoints detection to characterize the coordinate information of 25 keypoints of the human body with highresolution results;GIoU is the discriminant condition for establishing the association between PPEs and keypoints,and the authenticity of the PPE usage specification is effectively inferred by calculating the matching score of each PPE and the corresponding body region boxes.The experimental results show that the fusion detection model proposed in this thesis can identify whether healthcare workers are normatively donning multiple PPE with higher precision(95.81%),recall(96.38%),and F1-score(96.09%)on the MSDCD compared with models generated by VGG16,ResNet18,and CSPDarknet53 network training,and its parametric number(2.87M)and size(6.4MB)are also more lightweight.The fusion detection model performs more reliably in complex scenarios than the single object detection method.(4)A PPE-donning normative detection system for healthcare workers.Based on the theoretical study of the above detection models,and according to the actual needs of medical infection prevention and control monitoring,we deploy the research results of the detection models in hardware detection equipment and design a prototype system for intelligent detection of personal protection safety of healthcare workers.The system has stable detection efficiency while alleviating the pressure of medical infection prevention and control,and also has certain reference significance for other high-risk scenarios of safety and protection monitoring.