近年来由于施工人员不戴安全帽而导致的安全事故频率持续上升，因而亟需加强安全管理。基于深度学习的安全帽检测成为了当下重要的研究方向，现有安全帽检测算法存在数据和场景比例不均衡、算法检测精度低和误检较多的问题。因而，本文以深度学习算法模型为基础理论，从数据均衡化和模型改进两方面进行研究，并设计开发了安全帽检测系统，将改进算法集成到系统中。主要研究内容如下：

      （1）针对安全帽场景适应性差、被识别物体尺度不均衡的问题，提出了一种基于目标检测的数据双重均衡化方法（Object Detection Method of Data Double Equalization，OD-MDDE）。首先基于公开数据集SHWD扩建不同场景的安全帽数据，而后对扩建后的数据进行清洗得到了场景丰富的施工现场安全帽数据集（Self-establish Construction Safety Helmet Dataset，S-CSHD）；在此基础上采用图像增强算法对数据集中出现的色彩空间不佳的场景进行改善，提升模型对色彩的适应性；最后采用改进的K-means++算法重新聚类S-CSHD数据集的先验框尺寸，缓解数据尺度分布不均衡对算法带来的影响。经过数据均衡化后共得到包含15037张有效图片的训练样本集合，所有类平均精度达到93.56%，较YOLOv5s6.1提高了4.89%，证明了所做数据均衡化方法的有效性，为后续的研究做了良好的铺垫。

       （2）针对安全帽检测算法精度低、误检现象多的问题，提出了一种基于特征增强的安全帽检测算法（Feature Enhancement Network Algorithm，FENA）。首先在网络的颈部和主干层中引入了不同注意力模块；其次引入了解耦头部的概念，分别对分类和位置回归任务求解；然后在双融合结构中加入了双向级联结构，利用上下两个级联支路直接进行信息交换和传递；最后将模型主干层替换为ShuffleNetv2并引入GSConv卷积对网络进行轻量化。使用本文均衡化后的数据集S-CSHD对算法进行实验验证。实验结果表明，相比YOLOv5s6.1平均精度提高了2.0%，平均召回率提高了1.2%；经过轻量化的FENA体积减少了38.1%，以平均精度损失0.5%的代价将检测速度提升了47.8%。

       （3）基于上述研究结果，设计并实现了一个安全帽检测系统。该系统以PyQt5为框架，集成了改进的算法模型，可对所选图片和视频进行检测并将结果同步呈现，实时显示GPU使用率、FPS、mAP等指标，将算法的执行结果直观展示。

     In recent years, the frequency of accidents caused by construction workers not wearing safety helmets at construction sites has continued to rise, making it necessary to strengthen safety management. Safety helmet detection based on deep learning has become an important research direction. However, existing safety helmet detection algorithms suffer from data and scene imbalances, low detection accuracy, and high false positives. This thesis proposes a safety helmet detection system based on a deep learning algorithm model, which is developed and integrated with an improved algorithm into the system. The main research contents are as follows:

     （1）Aiming at the problems of poor adaptability of helmet scene and uneven scale of recognized objects, an object detection method based on data double equalization (OD-MDDE) is proposed. Firstly, the helmet data of different scenes are expanded based on the public data set SHWD, and then the expanded data are cleaned to obtain the scene-rich helmet data set (S-CSHD). On this basis, the image enhancement algorithm is used to improve the scenes with poor color space in the data set and improve the adaptability of the model to color. Finally, the improved K-means++ algorithm is used to recluster the prior frame size of S-CSHD dataset to alleviate the influence of uneven data scale distribution on the algorithm. After data equalization, a training sample set S-CSHD containing 15,037 valid pictures was obtained, and the average accuracy of all classes reached 93.56%, which was 4.89% higher than YOLOv5s6.1, which proved the effectiveness of the data equalization method and laid a good foundation for the follow-up research.

      （2）A feature enhancement network algorithm (FENA) based on feature enhancement is proposed to improve the accuracy and reduce false positives of safety helmet detection algorithms. The algorithm introduces different attention modules in the bottleneck and main layers of the network. The concept of decoupled head is also introduced to solve classification and position regression tasks separately. Moreover, the bi-directional concatenation structure is added to the double fusion structure for information exchange and transmission. Finally, the ShuffleNetv2 is used to replace the main branch of the model, and the GSConv convolution is introduced to lighten the network. Experimental results on S-CSHD show that the proposed algorithm achieves a 2.0% increase in average precision and 1.2% increase in average recall compared to YOLOv5s6.1, and the volume of the lightened model is reduced by 38.1%, while FENA speed is increased by 47.8% at the expense of 0.5% average precision loss.

     （3）Based on the above research, a safety helmet detection system is designed and implemented by integrating the improved algorithm model using PyQt5 as the framework. The system can perform detection on selected images and videos and present the results synchronously, providing real-time monitoring of GPU usage, FPS, mAP, and other indicators, making the algorithm's execution results intuitively displayed.

[1]常欣,刘鑫萌 .建筑施工人员不合理佩戴安全帽事故树分析[J].吉林建筑大学学报,2018,35(6):65-69.

[2]中华人民共和国住房和城乡建设部.2020年12月房屋市政工程生产安全事故情况通报[EB/OL].(2022-10-27).https://www.mohurd.gov.cn/gongkai/zhengce/zhengcefilelib/20

2210/20221026_768565.html

[3]郝忠,魏延晓.448例建筑施工工伤事故的标准统计特征分析[J].中国标准化,2017(02):245-247+249.

[4]周玉立,陈建国,袁宏永等.我国较大及以上建设工程施工事故特征分析[J].建筑安全,2021,36(02):46-52.

[5]边星,晋良海,陈雁高,郑霞忠,陈述,尹洁.施工作业人员佩戴安全帽行为意向研究[J].中国安全科学学报,2016,26(11):43-48.

[6]马舒琪,蔡嗣经,时训先.冲击载荷作用下工业安全帽防护性能研究[J].中国安科学学报,2014,24(10):133-137.

[7]Hale A R, Heming B H J, Carthey J, et al. Modelling of safety management systems[J]. Safety Science, 1997, 26(1-2): 121-140.

[8]姚群力,胡显,雷宏.深度卷积神经网络在目标检测中的研究进展[J].计算机工程与应用,2018,54(17):1-9.

[9]Girshick R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2014: 580-587.

[10]Girshick R. Fast r-cnn[C]//Proceedings of the IEEE international conference on computer vision. 2015: 1440-1448.

[11]Ren S, He K, Girshick R, et al. Faster r-cnn: Towards real-time object detection with region proposal networks[J]. Advances in neural information processing systems, 2015, 28.

[12]He K, Gkioxari G, Dollár P, et al. Mask r-cnn[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2961-2969.

[13]Redmon J, Farhadi A. Yolov3: An incremental improvement[J]. 2018.

[14]Bochkovskiy A, Wang C Y, Liao H Y M. Yolov4: Optimal speed and accuracy of object detection[J]. 2020.

[15]Li C, Li L, Jiang H, et al. YOLOv6: A single-stage object detection framework for industrial applications[J]. 2022.

[16]Oksuz K, Cam B C, Kalkan S, et al. Imbalance problems in object detection: A review[J]. IEEE transactions on pattern analysis and machine intelligence, 2020, 43(10): 3388-3415.

[17]李先飞,高琦,张乐.基于改进混合采样均衡化的质量规则获取[J].组合机床与自动化加工技术,2021,No.572(10):42-46.

[18]杨俊豪,李东升,陈春晓等.基于生成对抗网络的乳腺癌组织病理图像样本均衡化[J].生物医学工程研究,2020,39(02):161-166.

[19]陆泉,何超,陈静等.基于两阶段迁移学习的多标签分类模型研究[J].数据分析与知识发现,2021,5(07):91-100.

[20]Li B, Yao Y, Tan J, et al. Equalized focal loss for dense long-tailed object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022: 6990-6999.

[21]Liu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector[C]//Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14. Springer International Publishing, 2016: 21-37.

[22]Shrivastava A, Gupta A, Girshick R. Training region-based object detectors with online hard example mining[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 761-769.

[23]Pang J, Chen K, Shi J, et al. Libra r-cnn: Towards balanced learning for object detection[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019: 821-830.

[24]Wen C Y. The safety helmet detection technology and its application to the surveillance system[J]. Journal of forensic sciences, 2004, 49(4): 770-780.

[25]蔡利梅. 基于视频的煤矿井下人员目标检测与跟踪研究[D].中国矿业大学,2010.

[26]刘晓慧,叶西宁.肤色检测和Hu矩在安全帽识别中的应用[J].华东理工大学学报(自然科学版),2014,40(03):365-370.

[27]刘云波,黄华.施工现场安全帽佩戴情况监控技术研究[J].电子科技,2015,28(04):69-72.

[28]Shrestha K, Shrestha P P, Bajracharya D, et al. Hard-hat detection for construction safety visualization[J]. Journal of Construction Engineering, 2015, 2015(1): 1-8.

[29]冯国臣,陈艳艳,陈宁,李鑫,宋程程.基于机器视觉的安全帽自动识别技术研究[J].机械设计与制造工程,2015,44(10):39-42.

[30]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.

[31]徐守坤,王雅如,顾玉宛,李宁,庄丽华,石林.基于改进Faster RCNN的安全帽佩戴检测研究[J].计算机应用研究,2020,37(03):901-905.

[32]Park S, Yoon S, Heo J. Image-based automatic detection of construction helmets using R-FCN and transfer learning[J]. KSCE Journal of Civil and Environmental Engineering Research, 2019, 39(3): 399-407.

[33]Wang H, Zhang X, Sun Y, et al. Research and application of artificial technology for substation environment surveillance system[C]//2019 IEEE 8th Joint International Information Technology and Artificial Intelligence Conference (ITAIC). IEEE, 2019: 901-905.

[34]Jin M, Lu B, Zhang J, et al. Video Streaming Helmet Detection Algorithm Based On Feature Map Fusion And Faster RCNN[C]//2021 International Conference on Electronic Information Engineering and Computer Science (EIECS). IEEE, 2021: 470-474.

[35]李明山,韩清鹏,张天宇,王道累.改进SSD的安全帽检测方法[J].计算机工程与应用,2021,57(08):192-197.

[36]刘光品,刘云鹏,王仁芳.基于改进RetinaNet模型的室内安全帽佩戴检测研究[J].浙江万里学院学报,2020,33(06):97-103.

[37]马小陆,王明明,王兵.YOLOv3在安全帽佩戴检测中的应用研究[J].河北工程大学学报(自然科学版),2020,37(04):78-86.

[38]金雨芳,吴祥,董辉,俞立,张文安.基于改进YOLO v4的安全帽佩戴检测算法[J].计算机科学,2021,48(11):268-275.

[39]Sun S, Zhao S, Zheng J. Intelligent Site Detection Based on Improved YOLO Algorithm[C]//2021 International Conference on Big Data Engineering and Education (BDEE). IEEE, 2021: 165-169.

[40]王玲敏,段军,辛立伟.引入注意力机制的YOLOv5安全帽佩戴检测方法[J].计算机工程与应用,2022,58(09):303-312.

[41]赵睿,刘辉,刘沛霖,雷音,李达.基于改进YOLOv5s的安全帽检测算法[J/OL].北京航空航天大学学报:1-16[2023-03-30].

[42]Lin T Y, Maire M, Belongie S, et al. Microsoft coco: Common objects in context[C]//Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. Springer International Publishing, 2014: 740-755.

[43]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.

[44]Chen J, Liu D, Luo B, et al. Residual objectness for imbalance reduction[J]. Pattern Recognition, 2022, 130: 108781.

[45]Kong T, Sun F, Yao A, et al. Ron: Reverse connection with objectness prior networks for object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 5936-5944.

[46]杨先凤,李小兰,贵红军.改进的自适应伽马变换图像增强算法仿真[J].计算机仿真,2020,37(05):241-245.

[47]Arthur D, Vassilvitskii S. K-means++ the advantages of careful seeding[C]//Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete algorithms. 2007: 1027-1035.

[48]GitHub.SafetyHelmetWearing-Dataset(安全帽佩戴检测数据集) [EB/OL].(2019-12-17). https://github.com/njvisionpower/Safety-Helmet-Wearing-Dataset

[49]Wu H, Zhao J. An intelligent vision-based approach for helmet identification for work safety[J]. Computers in Industry, 2018, 100: 267-277.

[50]李鹏. 基于目标检测与深度估计的施工现场安全预警关键技术研究与实现[D].电子科技大学,2021.

[51]赵江河,王海瑞,吴蕾.FPN-CenterNet安全帽佩戴检测算法[J].计算机工程与应用,2022,58(14):114-120.

[52]Wang Q, Wu B, Zhu P, et al. ECA-Net: Efficient channel attention for deep convolutional neural networks[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 11534-11542.

[53]Ge Z, Liu S, Wang F, et al. Yolox: Exceeding yolo series in 2021[J]. 2021.

[54]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.

[55]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.

[56]Law H, Teng Y, Russakovsky O, et al. Cornernet-lite: Efficient keypoint based object detection[J]. 2019.

[57]Li H, Li J, Wei H, et al. Slim-neck by GSConv: A better design paradigm of detector architectures for autonomous vehicles[J]. 2022.

[58]Iandola F N, Han S, Moskewicz M W, et al. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size[J]. 2016.

[59]Ma N, Zhang X, Zheng H T, et al. Shufflenet v2: Practical guidelines for efficient cnn architecture design[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 116-131.

[60]Hou Q, Zhou D, Feng J. Coordinate attention for efficient mobile network design[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021: 13713-13722.

[61]Li X, Wang W, Hu X, et al. Selective kernel networks[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019: 510-519.

[62]李家旺. 基于卷积神经网络的雾天道路目标检测算法研究[D].西安电子科技大学,2020.

[63]Zhang Q L, Yang Y B. Sa-net: Shuffle attention for deep convolutional neural networks[C]//ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021: 2235-2239.

[64]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.