【作者】 王敏；

【导师】 邓宏涛；

【作者基本信息】 江汉大学 ， 计算机技术（专业学位）， 2025， 硕士

【摘要】 在建筑工地的高危工作环境中,安全帽是施工人员保护头部和减少伤害的必备防护装备。随着计算机视觉和深度学习技术的发展,目标检测技术在安全帽检测领域的研究具有广泛的价值,但在实际应用中仍面临着诸多挑战。由于检测的图像尺寸较小,极易导致检测过程中出现漏检和错检的问题;施工现场复杂多变的环境严重影响检测速度与精度;模型的参数量庞大和计算量复杂降低了实时监控系统的响应速度。针对这些问题,本文提出了一系列的优化策略。(1)针对现有研究模型中漏检和检测精度不足的问题,本文提出了一种改进的SICN-YOLOv7模型算法。在YOLOv7的主干网络中集合了SimAM模块,可以自动预测特征图中每个位置的重要性,强化对关键特征的关注,减少了漏检和误检的现象;引入Coord Conv卷积,直接将额外坐标的输入通道添加到卷积层中,增强模型处理坐标信息的能力,解决模型检测精度不足的问题;基于Normalized Wasserstein Distance方法优化了损失函数,该方法对导出的高斯分布进行相似性度量,加速了模型训练过程中的收敛速度。实验结果表明,这三方面改进措施使得模型的mAP值达到了93.8%,相比原模型提高了1.8%,有效增强了安全帽检测算法的准确性。(2)针对模型计算复杂和参数量大的问题,本文提出了一种改进的Tiny-YOLOv7模型算法。在YOLOv7的颈部网络采用RepConv模块作为主要计算单元,结合多尺度信息提取和特征复用的思想,进一步降低了模型的内存占用和计算量。将改进后的模块命名为Tiny-RepConv;使用轻量级上采样算子CARAFE,可以在不同位置自适应重组内核,有助于保持高分辨率特征图;最后整合Dynamic Head机制,统一目标检测中的尺度感知、空间感知和任务感知,提高模型在不同尺度和复杂场景下的检测性能。实验结果表明,这三方面改进措施使得模型的参数量减少至19.3M,mAP值仅降低1.0%,提出的方法在减少模型参数量的同时保持检测精度,降低了训练和部署难度。(3)本文基于上述两种算法设计了一套实时安全帽监测系统,该系统可以对施工工人的安全帽佩戴情况进行高效的实时监测和智能化管理。经过测试和验证,系统有效克服了传统检测系统的局限性,提升了系统检测的准确性和响应速度,满足实际工业中安全监控的需求。更多还原

【Abstract】 In the high-risk working environment of construction sites,safety helmets are essential protective equipment for construction workers to protect their heads and reduce injuries.With the development of computer vision and deep learning technology,the research of target detection technology in the field of safety helmet detection has a wide range of value,but it still faces many challenges in practical application.Due to the small size of the detected image,it is very easy to lead to the problem of missed and wrong detection in the detection process;the complex and changing environment of the construction site seriously affects the detection speed and accuracy;the large number of parameters of the model and the complexity of the computation reduces the response speed of the real-time monitoring system.Aiming at these problems,this paper proposes a series of optimisation strategies.(1)Aiming at the problems of leakage and insufficient detection accuracy in the existing research models,this paper proposes an improved algorithm for the SICN-YOLOv7 model.The SimAM module is assembled in the backbone network of YOLOv7,which can automatically predict the importance of each position in the feature map,strengthen the focus on key features,and reduce the phenomenon of leakage and misdetection;the Coord Conv convolution is introduced,which directly adds the input channels of additional coordinates to the convolution layer,enhancing the model’s ability to process the coordinate information,and solving the problem of insufficient detection accuracy of the model;based on the Normalized Wasserstein Distance method optimizes the loss function,which performs a similarity measure on the derived Gaussian distribution and accelerates the convergence speed during model training.The experimental results show that these three improvements make the model’s mAP value reach 93.8%,which is 1.8%higher than the original model,effectively enhancing the accuracy of the helmet detection algorithm.(2)Aiming at the problems of complex model computation and large number of parameters,an improved Tiny-YOLOv7 model algorithm is proposed in this paper.The RepConv module is adopted as the main computational unit in the neck network of YOLOv7,which combines the ideas of multi-scale information extraction and feature reuse to further reduce the memory occupation and computational volume of the model.The improved module is named Tiny-RepConv;the use of lightweight upsampling operator CARAFE,which can adaptively reorganise the kernel at different locations,helps to maintain the high-resolution feature maps;and finally the integration of the Dynamic Head mechanism,which unifies the scale-awareness,spatial-awareness,and task-awareness in the target detection,to improve the detection performance of the model in different scales and complex scenes.The experimental results show that these three improvements reduce the parameter volume of the model to19.3M,and the mAP value is only reduced by 1.0%.The proposed method reduces the number of model parameters while maintaining the detection accuracy,and reduces the difficulty of training and deployment.(3)In this paper,a real-time helmet monitoring system is designed based on the above two algorithms,which can carry out efficient real-time monitoring and intelligent management of construction workers’helmet wearing.After testing and verification,the system effectively overcomes the limitations of the traditional detection system,improves the accuracy and response speed of the system detection,and meets the needs of safety monitoring in the actual industry.更多还原