基于YOLOv5的作业人员违规穿戴手套情况检测

Detection of Non-compliant Glove Usage by Operators Based on YOLOv5

摘要:

针对作业人员违规穿戴手套情况检测问题，首先，采用高精度YOLOv5作为目标检测框架，并对其骨干网络进行了修改以提高其小目标识别能力；然后，在其强大的小目标识别能力的基础上增加了注意力机制（视觉Transformer）模块以提高整体识别精度，同时替换了原始的损失函数以进一步提高识别速度和准确率；最后，在采集的作业人员施工数据集上进行训练验证。实验结果表明，与原网络相比所提出的优化YOLOv5结构在验证数据集上的准确率显著提高，平均识别准确率能达到95%。

Abstract:

Aiming at the detection of non-compliant glove usage by operators, the high-precision YOLOv5 is adopted as the target detection framework and its backbone network is improved to enhance its ability of small-object recognition. Based on its strong small-object recognition capabilities, the attention mechanism (Visual Transformer) modules is incorporated to improve overall recognition accuracy. Additionally, the original loss function is replaced to further enhance recognition speed and accuracy. Finally, a data set collected from operators is trained and validated. Experimental results show that compared to the original network, the proposed optimized YOLOv5 structure has a significant improvement in accuracy, whose average recognition accuracy reaches 95% on the validation dataset.