基于改进YOLOv5的安全帽佩戴检测算法

Based on the improved YOLOv5 helmet wear detection algorithm

针对现有安全帽佩戴检测算法在复杂场景下存在密集目标检测难度大、小目标误检和漏检等问题,提出一种基于改进YOLOv5的安全帽佩戴检测算法。该算法主要在以下三个方面进行优化:通过在主干网络添加卷积块注意力模块(CBAM)来提取多个尺度的全局特征信息,使模型在通道和空间上更关注主要信息,得到更丰富的高层语义信息;将特征融合网络中的路径聚合网络(PAN)改进为加权双向特征金字塔网络(BiFPN),实现特征信息双向跨尺度连接和加权融合;将边界框回归损失函数改进为EIOU损失函数,加快边界框收敛速度和提高目标识别准确率。在自制的安全帽佩戴检测数据集上进行实验验证的结果表明:改进后的算法平均准确率(mAP)达到92.8%,相较于YOLOv5算法,改进后的算法在目标检测精确度和召回率上分别提升2.4%和1.8%。

Aiming at the problems of dense target detection difficulty, small target false detection and missed detection in complex scenarios, a helmet wearing detection algorithm based on improved YOLOv5 is proposed. The algorithm is mainly optimized in the following three aspects: by adding a convolutional block attention module to the backbone network to extract global feature information of multiple scales, so that the model pays more attention to the main information in channel and space, and obtains richer high-level semantic information;The path aggregation network in the feature fusion network is improved to a bidirectional feature pyramid network to achieve two-way cross-scale connection and weighted fusion of feature information;The bounding box regression loss function is improved to the EIOU loss function to speed up the bounding box convergence speed and improve the target recognition accuracy. The results of experimental verification on the self-made helmet wearing detection dataset show that the mean average precision of the improved algorithm reaches 92.8%, compared with the YOLOv5 algorithm,the improved algorithm improves the target detection accuracy and recall rate by 2.4% and 1.8%, respectively.