基于改进YOLOv4的安全帽佩戴检测方法

Helmet wearing detection method based on improved YOLOv4

针对建筑施工场地场景下远距离小目标安全帽佩戴检测问题,提出的一种改进YOLOv4的安全帽检测方法。将BN层和卷积层合并减少修改后的网络前向推理计算量,利用K-means聚类算法改进先验框维度,采用柔性NMS算法进行置信度权重修改解决标签重写问题,应用多尺度特征融合提升模型识别准确率。实验结果表明,该方法在安全帽数据集的检测任务中mAP提升2.91%;对低于32*32尺寸目标AP值相较于原算法提升6.02%,能够有效提升安全帽佩戴检测范围和准确率。

Aiming at the problem of helmet wearing detection of long-distance small targets in the scene of construction site, an improved helmet detection method of YOLOv4 was studied and proposed. The BN layer and convolution layer were combined to reduce the amount of forward reasoning calculation of the modified network, the prior frame dimension was improved using K-means clustering algorithm, and the confidence weight was modified using the flexible NMS algorithm to solve the problem of label rewriting, multi-scale feature fusion was used to improve the accuracy of model recognition. Experimental data show that the proposed method improves the map by 2.91% in the detection task of helmet data set. Compared with the original algorithm, the target AP value lower than 32*32 is increased by 6.02%, which can effectively improve the detection range and accuracy of helmet wearing.