个人防护装备检测以实时且准确地检测施工人员安全帽、安全带及反光衣等装备的规范化佩戴为目标，对防范事故发生具有重要意义。为了对施工人员佩戴的多类安全防护装备进行联合检测，同时改善复杂网络在资源受限的边缘设备上无法兼顾实时性和检测精度的问题，本文研究了基于改进YOLO的个人防护装备联合检测以及面向嵌入式设备的轻量化个人防护装备应用的联合检测方法。

    针对施工人员多类安全防护装备联合检测的问题，本文对YOLOv4算法的类别概率激活函数和非极大值抑制策略进行改进，设计了一种高精度端到端的个人防护装备联合检测算法YOLOv4-PPE。针对YOLOv4-PPE算法参数量大无法在嵌入式设备上实时检测的问题，设计了Ghost-Dw-PPE和CLSlim-PPE两种模型轻量化方法。第一种方法是对YOLOv4-PPE模型结构进行重构，首先用Ghost Bottleneck构成主干特征提取网络，其次在每个检测头选取合适位置插入SPP模块，最后重新设计特征融合结构的卷积模块和下采样操作。第二种方法是设计一种基于BN层缩放因子的通道剪枝与层剪枝方法（CLSlim），该方法对卷积模块的BN层缩放因子施加L1正则梯度进行稀疏化训练，通过全局剪枝阈值和局部安全阈值剔除大量冗余通道压缩模型参数量；通过层剪枝阈值修剪网络层提高模型检测速度。对YOLOv4-PPE、YOLOv4-Tiny-PPE分别进行CLSlim改进，实验结果表明：CLSlim-YOLOv4-PPE模型体积减少至4.15MB，mAP降低2.1%；CLSlim-YOLOv4-Tiny-PPE相较原模型各方面都有提升，其中模型体积为5.92MB，mAP较原模型提升0.8%；而GhostNet-Dw-PPE模型体积为44.6MB，mAP较原模型降低2.42%。对比本文设计的两类模型轻量化方法，CLSlim方法对模型压缩更加高效。

    最后选择CLSlim-YOLOv4-Tiny-PPE算法在RK3399pro为主处理器的嵌入式设备上进行应用测试，结果显示：mAP为92.3%，在嵌入式设备上的检测速度为0.0307秒，帧率约为33FPS，满足实际应用中每秒25帧的实时检测需求。因此，本文所设计的个人防护装备联合检测算法达到了实时性更高、识别精度更好的效果，具有一定的参考价值。

    Personal protective equipment (PPE) detection is aimed at real-time and accurate detection, standardized wearing of construction personnel's safety helmet, safety belt and reflective clothing, which is of great significance to prevent accidents. In order to combined detection of multiple types of safety protective equipment worn by the construction personnel, and improve the problem that complex network can not give consideration to both real-time and detection accuracy on resource-limited edge devices, this paper studies the combined detection method of PPE applications based on improved YOLO algorithm, and lightweight PPE applications for embedded device.

    To solve the problem that combined detection of multiple safety protective equipment, A high-precision and end-to-end PPE combined detection algorithm, YOLOv4-PPE is designed by improving the class probability activation function, and non-maximum suppression strategy of YOLOv4 algorithm. To solve the problem, which the YOLOv4-PPE has too many parameters, and cannot be detected in real time on embedded devices, two methods of model lightweight were designed: Ghost-Dw-PPE and CLSlim-PPE. The first method was to reconstruct the YOLOv4-PPE model structure. First, Ghost Bottleneck was used to form the backbone feature extraction network, and then the Spatial Pyramid Pooling (SPP) module was inserted into the appropriate position of each detection head. Finally, the convolution module and down-sampling operation of feature fusion structure were redesigned. The second method was to design a channel pruning and layer pruning method (CLSlim), based on BN layer scaling factor. This method applies L1 regularization and gradient sparse training, on the scaling factor of Batch Normalization (BN) layer in the convolution module. A large number of redundant channel compression model parameters were removed by global pruning threshold, and local safety threshold. layer pruning threshold were used to improve inference speed. The CLSlim lightweight be used to improves YOLOv4-PPE and YOLOv4-Tiny-PPE model separately. The results show that the volume of CLSlim-YOLOv4-PPE model is reduced to 4.15MB and mAP decreases by 2.1%; CLSlim-YOLOV4-Tiny-PPE improves in all aspects compared with the original model, among which the model volume is 5.92MB and mAP is 0.8% higher than the original model. However, the volume of Ghost-Dw-PPE model is 44.6MB, and the mAP is reduced by 2.42% compared with the original model. Compared with the two types of model lightweight methods designed in this paper, CLSlim method is more efficient for model compression.

    Finally, the CLSlim-YOLOv4-Tiny-PPE algorithm was selected for application test on embedded devices with RK3399pro as main processor. The results show that: mAP is 92.3%, the detection speed on the embedded device is 0.0307 seconds, and the frame rate is about 33FPS, which meets the real-time detection requirements of 25 frames per second in practical application. Therefore, the combined detection algorithm of personal protective equipment designed in this paper achieves higher real-time performance and better identification accuracy, which has certain reference value.

[1] 王亚光. 基于安全生产对建筑施工的管理预防[J]. 现代商贸工业, 2022, 43(01): 186-188.

[2] 张泾杰, 韩豫, 姚佳玥, 等. 建筑工人安全装备自动检查系统设计及实现[J]. 施工技 术, 2017, 46(24): 4.

[3] 韩泽佳, 肖秦琨, 张立旗. 基于改进SSD安全头盔反光衣检测算法[J]. 自动化与仪表, 2021, 36(9): 6.

[4] Wu J, Cai N, Chen W, et al. Automatic detection of hardhats worn by construction personnel: A deep learning approach and benchmark dataset[J]. Automation in Construction, 2019, 106: 102894: 1-7.

[5] Ren S, He K, Girshick R, et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 39(6): 1137-1149.

[6] Redmon J, Farhadi A. YOLO9000: better, faster, stronger[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 7263-7271.

[7] Lin T Y, Goyal P, Girshick R, et al. Focal Loss for Dense Object Detection[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, PP(99): 2999-3007.

[8] Zhao Q, Sheng T, Wang Y, et al. M2Det: A Single-Shot Object Detector Based on Multi-Level Feature Pyramid Network[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2019, 33: 9259-9266.

[9] Zhang S, Wen L, Bian X, et al. Single-shot refinement neural network for object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 4203-4212.

[10] Bochkovskiy A, Wang C Y, Liao H. YOLOv4: Optimal Speed and Accuracy of Object Detection[EB/OL]. http://arxiv.org/abs/2004.10934, 2020.

[11] Howard A G, Zhu M, Chen B, et al. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications[EB/OL]. http://arxiv.org/abs/1704.04861, 2017.

[12] Zhang X, Zhou X, Lin M, et al. ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices[EB/OL]. http://arxiv.org/abs/1707.01083, 2017.

[13] 符惠桐, 王鹏, 李晓艳, 等. 面向移动目标识别的轻量化网络模型[J]. 西安交通大学 学报, 2021, 55(7): 124-131

[14] 王燕妮, 孙雪松,余丽仙. 增强感受野的轻量化合成孔径雷达船舶检测算法[J]. 光子 学报, 2022, 2(51): 266-278.

[15] 陈科圻, 朱志亮, 邓小明等. 多尺度目标检测的深度学习研究综述[J]. 软件学报, 2021, 32(4): 1201-1227.

[16] Lin T Y, Dollár P, Girshick R, et al. Feature pyramid networks for object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 2117-2125.

[17] Liu S, Qi L, Qin H, et al. Path aggregation network for instance segmentation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 8759-8768.

[18] Hu J, Shen L, Sun G. Squeeze-and-excitation networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 7132-7141.

[19] Liu S, Huang D, Wang Y. Learning Spatial Fusion for Single-Shot Object Detection[EB/OL]. http://arxiv.org/abs/1911.09516, 2019.

[20] Yu G, C Q, L W, et al. PP-PicoDet: A Better Real-Time Object Detector on Mobile Devices[EB/OL]. http://arxiv.org/abs/2111.00902, 2021.

[21] Dog-qiuqiu. GitHub-dog-qiuqiu/Yolo-Fastest: Based on yolo's ultra-lightweight universal target detection algorithm, the calculation amount is only 250mflops, the ncnn model size is only 666kb, the Raspberry Pi 3b can run up to 15fps+, and the mobile terminal can run up to 178fps+[EB/OL]. https://github.com/dog-qiuqiu/Yolo-Fastest, 2022-2-17

[22] Cheng Y, Wang D, Zhou P, et al. A Survey of Model Compression and Acceleration for Deep Neural Networks[EB/OL]. http://arxiv.org/abs/1710.09282, 2017.

[23] Howard A G, Zhu M, Chen B, et al. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications[J]. CoRR abs/1704.04861, 2017.

[24] Sandler M, Howard A, Zhu M, et al. Inverted Residuals and Linear Bottlenecks: Mobile Networks for Classification, Detection and Segmentation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2018: 1580-1589

[25] Howard A, Sandler M, Chu G, et al. Searching for mobilenetv3[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019: 1314-1324..

[26] Lee Y, Park J. Centermask: Real-time anchor-free instance segmentation[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 13906-13915..

[27] Tan M, Pang R, Le Q V. Efficientdet: Scalable and efficient object detection[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 10781-10790.

[28] Han K, Wang Y, Tian Q, et al. Ghostnet: More features from cheap operations[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 1580-1589.

[29] He Y, Lin J, Liu Z, et al. Amc: Automl for model compression and acceleration on mobile devices[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 784-800.

[30] Elsken T, Metzen J H, Hutter F. Neural architecture search: A survey[J]. The Journal of Machine Learning Research, 2019, 20(1): 1997-2017.

[31] 王兵, 李文璟, 唐欢. 改进YOLO v3算法及其在安全帽检测中的应用[J]. 计算机工 程与应用, 2020, 56(09): 33-40

[32] 吴冬梅, 王慧, 李佳. 基于改进Faster RCNN的安全帽检测及身份识别[J]. 信息技术 与信息化, 2020(01): 17-20.

[33] 刘学锋王秋茗孙广玲陆小锋钱国. 智慧工地中低分辨率的安全帽状态识别[J]. 电子 测量技术, 2020, 43(15): 63-67.

[34] Han G, Zhu M, Zhao X, et al. Method based on the cross-layer attention mechanism and multiscale perception for safety helmet-wearing detection[J]. Computers & Electrical Engineering, 2021, 95: 107458.

[35] 韩豫, 张泾杰, 孙昊, 等. 基于图像识别的建筑工人智能安全检查系统设计与实现[J]. 中国安全生产科学技术, 2016, 12(10): 142-148.

[36] 罗小安. 智慧工地基于视频的高空作业人员安全带穿戴检测方法[P]. 湖北省： CN109635758B, 2021-07-09.

[37] 武汉倍特威视系统有限公司. 智慧工地安全带识别系统[EB/OL]. http://ai.betvsys.com/htm/318.html, 2022-2-19

[38] 刘欣宜, 张宝峰, 符烨等. 基于深度学习的污染场地作业人员着装规范性检测[J]. 中 国安全生产科学技术, 2020, 16(07): 169-175.

[39] 莫蓓蓓, 吴克河. 引入Self-Attention的电力作业违规穿戴智能检测技术研究[J]. 计 算机与现代化, 2020(02): 115-121+126.

[40] 张春堂, 管利聪. 基于SSD-MobileNet的矿工安保穿戴设备检测系统[J]. 工矿自动化, 2019, 45(06): 96-100.

[41] Xie Z, Liu H, Li Z, et al. A convolutional neural network based approach towards real-time hard hat detection[C]//2018 IEEE International Conference on Progress in Informatics and Computing (PIC). IEEE, 2018: 430-434.

[42] Nath N D, Behzadan A H, Paal S G. Deep learning for site safety: Real-time detection of personal protective equipment[J]. Automation in Construction, 2020, 112: 103085: 1-20.

[43] Wang C Y, Liao H Y M, Wu Y H, et al. CSPNet: A new backbone that can enhance learning capability of CNN[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. 2020: 390-391.

[44] Purkait P, Zhao C, Zach C. SPP-Net: Deep absolute pose regression with synthetic viewshttp://arxiv.org/abs/1712.03452, 2017.

[45] 朱彦. 对高空作业坠落防护装备及其技术措施的研究[J]. 中国个体防护装备, 2015(06): 44-18.