Detection Method for Laboratory PPE Compliance Wearing Based on Human Key Points

doi:10.16182/j.issn1004731x.joss.25-0415

Abstract

Abstract:

To address the problems of high missed detection rate and inaccurate judgment of wearing compliance when multi-scale and multi-category targets of laboratory personnel's safety protective equipment are detected in a complex laboratory environment, this paper proposesa laboratory personnel's standard personal protective equipment (PPE) wearing detection method (multi-scale multi-target joint key point detection method, MSMT-JKDM) that integrates multi-scale features and human key points. The multi-scale adaptive down sampling (MSA-Down) module and the cascaded group attention transformer (CGA Former) are introduced to enhance the feature representation ability of PPE (especially small targets such as goggles and gloves) in laboratory detection scenarios, effectively alleviating the problems of small targets being easily ignored and feature confusion. To solve the problem that directly fusing multi-scale features of different types of PPEs in complex environments may lead to feature redundancy, spatial misalignment, and inconsistent object representation, the selective boundary-aware multi-scale recalibration feature pyramid network (SBMSRFPN) is adopted to improve the representation ability of cross-scale features through multi-scale fusion, boundary recalibration, and attention guidance. An adaptive anchor-free detection head (AAF-Head) is designed to improve the positioning accuracy and detection efficiency of multi-scale targets. The relevant coordinate information obtained from the two object branches of protective equipment and human body parts is jointly inferred to accurately judge the standardization of PPE wearing of the experimenters. Experimental verification is carried out on this dataset. The results show that the mAP50, mAP50-95, and standard wearability index standardized wearing accuracy (SWA) are improved by 3.7%, 6.2%, and 5%, respectively compared with those of the baseline model. The mAP50 reaches 82.8% and 87.9% on the public datasets GDUT-HWD and SHWD, respectively, better than that of the existing mainstream detection methods, further verifying the effectiveness of the model.

Key words: laboratory personnel, PPE wearing regulations, MSMT-JKDM, multi-scale multi-target detection, key point detection

CLC Number:

TP391.9

Peng Lijun, Su Tingqi, Liu Peijin, He Lin, Zhou Xiewu, Zhang Minxin. Detection Method for Laboratory PPE Compliance Wearing Based on Human Key Points[J]. Journal of System Simulation, 2026, 38(5): 1365-1382.

Figures/Tables 19

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Fig. 5

Fig. 6

Fig. 7

Fig. 8

Fig. 9

Fig. 10

Table 1

Fig. 11

Table 2

Comparison results of different algorithms on DCDLE dataset

模型	P/%	R/%	mAP50/%	$m A P$ 50~95/%	计算量×10⁹	参数量×10⁶	SWA/%
YOLOv5n	91.4	81.5	86.3	57.7	7.1	25.0	71.7
YOLOv6n	90.3	85.1	87.5	61.0	11.8	42.3	85.0
YOLOv8n	91.0	83.7	87.7	60.2	8.1	30.1	68.3
YOLOv9t	91.6	82.7	87.3	60.0	7.6	19.7	96.7
YOLOv10n	78.5	71.9	79.6	53.8	8.2	26.9	75.0
YOLO11n	88.0	84.1	86.7	58.0	6.3	25.8	93.3
YOLOv12n	88.7	81.5	85.6	57.9	5.8	25.1	75.0
Hyper-YOLO	90.3	84.3	88.2	59.5	9.5	36.2	70.0
RT-DETRl	71.0	75.1	78.3	54.2	100.6	284.6	95.0
RT-DERT-resnet50	78.6	80.3	85.0	60.3	125.7	419.5	86.7
Ours	94.0	85.1	89.9	61.6	10.6	28.7	98.3

Table 2

Table 3

Table 4

Table 5

Table 6

Fig. 12

Fig. 13

References 23

[1]	教育部. 关于加强高校实验室安全工作的意见[J]. 中华人民共和国教育部公报, 2019(5): 29-31.
[2]	Blacksell S D, Dhawan Sandhya, Kusumoto Marina, et al. Laboratory-acquired Infections and Pathogen Escapes Worldwide Between 2000 and 2021: A Scoping Review[J]. The Lancet. Microbe, 2024, 5(2): e194-e202.
[3]	王忠玉. 智能视频监控下的安全帽佩戴检测系统的设计与实现[D]. 北京: 北京邮电大学, 2018.
	Wang Zhongyu. Design and Implementation of Detection System of Wearing Helmets Based on Intelligent Video Surveillance[D]. Beijing: Beijing University of Posts and Telecommunications, 2018.
[4]	李春亚, 王建华, 尹土兵, 等. 个人防护装备穿戴安全检测技术的进展与应用[J]. 武汉理工大学学报, 2024, 46(10): 149-155, 162.
	Li Chunya, Wang Jianhua, Yin Tubing, et al. Advancements and Applications of Wearable Safety Detection Technology for Personal Protective Equipment[J]. Journal of Wuhan University of Technology, 2024, 46(10): 149-155, 162.
[5]	Jiang Huaizu, Learned-Miller E. Face Detection with the Faster R-CNN[C]//2017 12th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2017). Piscataway: IEEE, 2017: 650-657.
[6]	Zhang Qiang, Pei Ziyu, Guo Rong, et al. An Automated Detection Approach of Protective Equipment Donning for Medical Staff Under COVID-19 Using Deep Learning[J]. CMES - Computer Modeling in Engineering and Sciences, 2022, 132(3): 845-863.
[7]	Chen Weiming, Li Changfan, Guo Hailin. A Lightweight Face-assisted Object Detection Model for Welding Helmet Use[J]. Expert Systems with Applications, 2023, 221: 119764.
[8]	Cao Zhe, Hidalgo G, Simon T, et al. OpenPose: Realtime Multi-person 2D Pose Estimation Using Part Affinity Fields[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43(1): 172-186.
[9]	Sun Ke, Xiao Bin, Liu Dong, et al. Deep High-resolution Representation Learning for Human Pose Estimation[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2019: 5686-5696.
[10]	Xiong Ruoxin, Tang Pingbo. Pose Guided Anchoring for Detecting Proper Use of Personal Protective Equipment[J]. Automation in Construction, 2021, 130: 103828.
[11]	Gong Faming, Ji Xiaofeng, Gong Wenjuan, et al. Deep Learning Based Protective Equipment Detection on Offshore Drilling Platform[J]. Symmetry, 2021, 13(6): 954.
[12]	Su Qin, Zhang Juling, Chen Mingju, et al. PW-YOLO-pose: A Novel Algorithm for Pose Estimation of Power Workers[J]. IEEE Access, 2024, 12: 116841-116860.
[13]	Yang Bin, Xiao Hongru, Zhang Binghan. A Unified Object and Keypoint Detection Framework for Personal Protective Equipment Use[J]. Developments in the Built Environment, 2024, 20: 100559.
[14]	Jamali Ali, Kumar Roy Swalpa, Bhattacharya Avik, et al. Local Window Attention Transformer for Polarimetric SAR Image Classification[J]. IEEE Geoscience and Remote Sensing Letters, 2023, 20: 1-5.
[15]	Liu Xinyu, Peng Houwen, Zheng Ningxin, et al. EfficientViT: Memory Efficient Vision Transformer with Cascaded Group Attention[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2023: 14420-14430.
[16]	Tang Feilong, Xu Zhongxing, Huang Qiming, et al. DuAT: Dual-aggregation Transformer Network for Medical Image Segmentation[C]//Chinese Conference on Pattern Recognition and Computer Vision. Singapore: Springer Nature Singapore, 2024: 343-356.
[17]	Lin T Y, Maire M, Belongie S, et al. Microsoft COCO: Common Objects in Context[C]//Computer Vision – ECCV 2014. Cham: Springer International Publishing, 2014: 740-755.
[18]	Wu Jixiu, Cai Nian, Chen Wenjie, et al. Automatic Detection of Hardhats Worn by Construction Personnel: A Deep Learning Approach and Benchmark Dataset[J]. Automation in Construction, 2019, 106: 102894.
[19]	Njvisionpower. Safety-helmet-wearing-dataset[EB/OL]. [2025-05-01]. .
[20]	Fang Qi, Li Heng, Luo Xiaochun, et al. Detecting Non-hardhat-use by a Deep Learning Method from Far-field Surveillance Videos[J]. Automation in Construction, 2018, 85: 1-9.
[21]	Yang Bin, Wang Jie. An Improved Helmet Detection Algorithm Based on YOLO V4[J]. International Journal of Foundations of Computer Science, 2022, 33(06N07): 887-902.
[22]	Zhou Fangbo, Zhao Huailin, Nie Zhen. Safety Helmet Detection Based on YOLOv5[C]//2021 IEEE International Conference on Power Electronics, Computer Applications (ICPECA). Piscataway: IEEE, 2021: 6-11.
[23]	Wei Lihong, Liu Panpan, Ren Haihui, et al. Research on Helmet Wearing Detection Method Based on Deep Learning[J]. Scientific Reports, 2024, 14(1): 7010.

关键点	Class 1 (正确穿戴)	Class 2 (未正确穿戴)	验证逻辑
Left eye	glass	unglass	护目镜是否覆盖左眼
Right eye	glass	unglass	护目镜是否覆盖右眼
Left shoulder	coat	uncoat	实验服是否覆盖左肩
Right shoulder	coat	uncoat	实验服是否覆盖右肩
Left hip	coat	uncoat	实验服是否覆盖左髋
Right hip	coat	uncoat	实验服是否覆盖右髋
Left wrist	glove	unglove/mismatch	手套是否覆盖左手腕
Right wrist	glove	unglove/mismatch	手套是否覆盖右手腕

模型	类别	P/%	R/%	mAP50/%	mAP50~95/%	P_C-SWA/%
YOLO11n	person	98.9	98.6	99.1	79.6
	coat	99.3	99.5	99.5	87.4	100
	glass	94.2	96.8	97.5	62.4	94.7
	glove	81.6	81.0	87.2	55.3	85.9
	uncoat	78.2	70.9	76.3	32.7
	unglass	98.5	96.1	97.6	71.6
	unglove	65.2	45.9	49.5	17.1
Ours	person	98.3	98.3	99.3	82.5
	coat	99.6	98.5	99.2	88.4	100
	glass	96.7	94.9	96.7	65.2	95.4
	glove	92.0	82.8	89.5	60.6	88.0
	uncoat	89.4	71.7	83.6	37.0
	unglass	98.9	97.0	97.7	77.2
	unglove	83.3	52.4	63.0	25.7

实验	MSA-Down	CGAFormer	SBMSRFPN	AFF-Head	mAP50/%	mAP50~95/%	P_SWA/%
1					86.7	58.0	93.3
2	√				87.9	60.3	92.5
3		√			88.7	59.4	93.6
4			√		88.9	60.3	94.7
5				√	87.2	58.4	95.7
6	√	√			88.4	59.9	96.2
7	√		√		89.7	61.0	97.2
8	√	√	√		88.8	61.2	97.9
9	√	√	√	√	89.9	61.6	98.3

模型	P/%	R/%	mAP50/%	mAP50~95/%	参数量×10⁶	计算量×10⁹
YOLOv5n	86.2	72.7	80.3	48.4	25.0	7.1
YOLOv8n	88.8	72.9	80.7	48.8	26.8	6.8
YOLOv9t	86.5	71.1	79.6	47.6	17.3	6.4
YOLOv10n	82.5	68.0	75.9	45.3	26.9	8.2
YOLO11n	85.4	74.6	80.9	48.4	25.8	6.3
YOLOv12n	86.9	72.8	80.0	47.1	25.1	5.8
RT-DERT-l	75.5	66.4	72.8	42.0	284.5	100.6
Hyper-YOLO	88.4	74.9	82.1	48.2	36.2	9.5
Ours	89.7	74.1	82.8	49.0	28.7	10.6

模型	P/%	R/%	mAP50/%	mAP50~95/%	参数量×10⁶	计算量×10⁹
YOLOv5n	88.7	80.4	87.2	51.9	25.3	7.1
YOLOv8n	88.9	80.1	87.9	52.7	30.1	8.1
YOLOv9t	88.8	79.8	86.9	52.1	19.7	7.6
YOLOv10n	84.7	76.4	84.1	49.8	26.9	8.2
YOLO11n	88.8	79.7	87.7	52.5	25.8	6.3
YOLOv12n	88.5	78.7	86.7	51.4	25.1	5.8
RTDERT-l	82.0	72.1	80.0	45.6	254.5	100.6
Hyper-YOLO	88.3	80.1	85.9	51.1	36.2	9.5
Improvement of Faster RCNN^[20]	82.2	79.2	80.3	54.4
Improvement of YOLOv4^[21]	88.4	74.0	79.4	44.1
Improvement of YOLOv5^[22]	85.3	75.2	81.2	52.6
BiFEL-YOLOv5s^[23]	86.5	77.9	82.8	57.7
Ours	90.2	80.8	87.9	59.0	28.7	10.6