基于 YOLO V5 的安全帽快速检测-《计算机技术与发展》

文章信息/Info

Author(s):: QIN Jian-yong* ; ZHU Ming-zeng; QIN Qiu-qin; MO Zi-ying; QIN Jing-tao; Hezhou Power Supply Bureau,Guangxi Power Grid Co. ,Ltd. ,Hezhou 542800,China

Keywords:: YOLO V5; safety helmet; deep learning; distance-IOU; Density Peak cluster

摘要:: 安全帽的佩戴检测可有效防范因未佩戴安全帽而造成的施工事故,在安全生产过程中发挥重要作用。该文提出一种基于 YOLO V5 的安全帽自动、高效检测方法。首先,基于 Density Peak 聚类算法对训练数据进行聚类分析,根据Density Peak 决策图优化 anchor 尺度。然后,综合 anchor 的位置和面积关系引入 DIOU 损失函数来提高模型的训练精度。最后,建立安全帽佩戴样本数据库,采用 YOLO V5 模型进行训练和测试。为验证算法的有效性,该方法在自建的安全帽佩戴数据集上进行了测试。实验结果表明,该方法的 mPA0. 5 可达 77. 2% ,图像推理速率可达 140 帧 / 秒,具有精度较高、检测速度快等优点。此外,该方法也可以有效检测多目标图像,能够满足复杂环境下安全帽的自动检测需求,且容易部署和推广到嵌入式应用设备中。

Abstract:: The detection of safety helmet plays an important role in the prevention of safety accidents,and we propose an automatic and efficient detection method of safety helmet. First,a Density Peak algorithm is used to analyze the size feature of safety helmet,which helps to increase the accuracy of detection. Then,according to area and distance of prediction anchor and truth anchor,a DIoU ( distance-IoU) function is introduced to build the loss function of YOLO v5. Finally,the deep learning framework YOLO v5 is used to train and test the safety helmet. Experimental results on the created dataset show the efficiency and effectiveness of the proposed method on safety helmet detection,which obtains mAP0. 5,image inference ratio with 77. 2% and 140 frame / s. In addition,the proposed method is also able to detect the images with multi-object, and can easily be applied for the embedded device.