基于改进的YOLOv4骑行人员佩戴头盔目标检测研究

Research on object detection of YOLOv4 cyclist wearing helmet based on improvement

在交通场景下,现有YOLOv4算法检测电动自行车及摩托车驾驶员头盔佩戴时存在漏检率偏高、定位准确度较差、中小目标易聚集等问题。对此,文章提出了一种改进的YOLOv4算法。首先,选用轻量级网络MobilenetV3作为主干特征网络,并将YOLOv4网络中尺寸为3×3的标准卷积层均替换为深度可分离卷积,缩减了模型参数量;其次,引入ECA注意力机制与深度可分离卷积结合,从而替换PANet模块中的深度可分离卷积并增强特征网络表现力;最后,使K-means++聚类算法与Intersection-over-Union(IoU)相结合,并与anchors比较聚类,增强了聚类效果。实验结果表明,改进的YOLOv4算法mAP值达到了97.85%,比YOLOv4算法高出1.22%,检测速度由33.55 fps提升至50.60 fps。在满足精确性的前提下,改进的YOLOv4算法在头盔佩带检测场景中具有较好的性能,更利于轻量化部署。

In traffic scenarios,the existing YOLOv4 algorithm has problems such as high missed detection rate,poor positioning accuracy,and easy clustering of small and medium-sized targets when detecting helmets worn by electric bicycle and motorcycle drivers.This article proposes an improved YOLOv4 algorithm for this.Firstly,the lightweight network MobilenetV3 was selected as the backbone feature network,and the standard convolutional layers with a size of 3×3 in the YOLOv4 network were replaced with depthwise separable convolutions,reducing the number of model parameters.Secondly,the ECA attention mechanism is combined with depthwise separable convolution to replace the depthwise separable convolution in the PANet module and enhance the performance of the feature network.Finally,the K-means++clustering algorithm was combined with Intersection-over-Union(IoU)and compared with anchors to improve clustering performance.The experimental results show that the improved YOLOv4 algorithm has an mAP value of 97.85%,which is 1.22%higher than the YOLOv4 algorithm.The detection speed has been improved from 33.55 fps to 50.60 fps.Under the premise of meeting accuracy,the improved YOLOv4 algorithm has good performance in helmet wearing detection scenarios,which is more conducive to lightweight deployment.