基于单阶段半监督目标检测的建筑工人检测算法  

Construction worker detection algorithm based on one-stage semi-supervised object detection

作　　者：方莉 赵志峰 严铮 戴振国 陈国栋[1] FANG Li;ZHAO Zhifeng;YAN Zheng;DAI Zhenguo;CHEN Guodong(College of Physics and Information Engineering,Fuzhou University,Fuzhou 350108,China)

机构地区：[1]福州大学物理与信息工程学院,福建福州350108

出　　处：《微电子学与计算机》2025年第2期20-30,共11页Microelectronics & Computer

基　　金：福建省科技计划引导性项目(2021H0013);福建省科技型中小企业创新资金项目(2021C0019);福州市科技计划(2022-XG-001)。

摘　　要：建筑工人目标检测对于提升建筑施工安全具有重要的应用价值。随着智慧工地的推广,施工区域的视频监控覆盖率不断增加,获取大量未标注的建筑工人图像变得更为便捷,而有标注数据图像依然稀缺而昂贵。半监督学习方法是解决有标注数据缺乏问题的有效办法。然而,施工环境中存在着环境混乱、目标遮挡以及监控画面可视度低等问题,导致半监督目标检测模型在伪标签生成阶段难以平衡数量与质量。已有的半监督目标检测算法大多基于两阶段目标检测模型设计,未能满足对建筑工人检测实时性的要求。为了解决上述问题,提出了一种针对施工场景设计的单阶段半监督建筑工人目标检测算法。首先,将半监督目标检测应用于建筑工人目标检测任务,有效解决了标注数据缺乏的问题。其次,提出软阈值优化方法,为低置信样本分配权重,从而扩充伪标签的数量。接着,引入图像信息熵概念来评估样本检测难度,并提出自适应阈值选择算法以根据样本难度调整伪标签的阈值,进而提高训练初期的伪标签质量。最后,通过增加残差特征金字塔网络和上下文增强模块提升对小目标的检测能力。实验证明,在自建的施工区域建筑工人检测数据集上,所提出的算法在解决单阶段半监督建筑工人目标检测问题方面表现出显著优势。The detection of construction workers holds significant practical value for enhancing construction site safety.With the proliferation of smart construction sites,the coverage of video surveillance in construction areas continues to expand,making it more convenient to obtain a substantial amount of unlabeled construction worker images.However,labeled data remains scarce and expensive.Semi-supervised learning methods prove to be effective in addressing the issue of insufficient labeled data.Nevertheless,environmental clutter,object occlusion,and low visibility in monitoring footage pose difficulties in achieving a balance between quality and quantity of pseudo-labels generated during the training phase of Semi-Supervised Object Detection(SSOD)model.Existing SSOD algorithms predominantly rely on two-stage model designs,failing to meet the real-time requirements for construction worker detection.To tackle these issues,a single-stage semi-supervised construction worker object detection algorithm tailored for construction scenes(the Balanced Teacher Model)is proposed.Firstly,SSOD is applied to the detection of construction workers,effectively mitigating the problem of limited labeled data.Secondly,Soft Threshold Optimization(STO)is introduced to assign weights to low-confidence samples,thereby augmenting the quantity of pseudo-labels.Subsequently,image information entropy is incorporated to assess detection difficulty for each picture.Adaptive Threshold Selection(ATS)is proposed to adjust pseudo-label thresholds based on detection difficulty,thereby improving the quality of pseudo-labels in the early stages of training.Finally,the detection capability for small targets is improved by introducing Residual Feature Pyramid Network(ReFPN)and Context Augmentation Module(CAM).Experimental results on a self-constructed dataset for construction worker detection demonstrate the significant advantages of the proposed model in addressing the single-stage semi-supervised construction worker detection challenges within the construc

关 键 词：建筑工人目标检测 半监督目标检测 数量质量权衡 单阶段目标检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]