技术与市场

为弥补人工巡检实时性的不足，提出一套基于YOLOv8n的工业危险行为检测系统。该系统利用公开个人防护装备(personal protective equipment, PPE)数据集，并以1 000张现场样本微调，融合压缩-激励模块(squeeze-and-excitation, SE)或卷积块注意力模块(convolutional block attention module, CBAM)与“Vote-Cool”10帧滑动投票-50帧冷却策略，在轻量前提下提升小目标和遮挡场景鲁棒性。试验结果表明，在综合PPE基准集上改进模型取得平均精度均值(mAP@0.5)为40.2%,人员/PPE检测精度为94.1%/86.3%,RTX3080单卡全流程推理165 FPS(每秒传输帧数)。结果验证了方案在无大规模私有数据集条件下也能满足工业安全监控的实时性与可靠性，为基于公开数据的检测系统的快速落地提供了可行路径。

To address the real-time limitations of manual inspection, this paper proposes an industrial hazardous behavior detection system based on YOLOv8n.The system utilizes public personal protective equipment(PPE) datasets and is fine-tuned with 1 000 on-site samples, integrating squeeze-and-excitation(SE)module or convolutional block attention module(CBAM) and a “Vote-Cool” strategy(10-frame sliding vote and 50-frame cool-down) to enhance robustness for small targets and occluded scenes while maintaining a lightweight structure.Experimental results show that the improved model achieves an mAP@0.5 of 40.2% on a comprehensive PPE datasets.The detection accuracy for persons and PPE reaches 94.1% and 86.3%,respectively.The full-pipeline inference speed on a single RTX3080 is 165 FPS.The experiment verifies that this system can meet the real-time and reliability requirements for industrial safety monitoring even without large-scale private datasets, which provides a feasible path for rapid deployment based on public data.