基于优化改进的YOLOv4算法在骑行头盔佩戴检测上的研究

Research on the Wearing Detection of Cycling Helmet Based on Optimized and Improved YOLOv4 Algorithm

针对当前算法在交通环境中对电动车和摩托车驾驶员的头盔配戴检测表现出较高的漏检率和定位精度不足等问题,本文提出了一种优化改进的YOLOv4算法。引入了轻量级网络MobileNetV2替代CSPDarknet53主干网络,不仅显著减少了参数数量和计算成本,而且保持了必要的表征能力。采用CA注意力机制结合深度可分离卷积提升网络特征的代表性,并且与深度可分离卷积结合进一步优化了计算效率。为了更有效地利用不同层次的特征并提升检测性能,引入了SPP+最大池化结构。在不同的空间尺度上捕捉和融合特征,促进模型捕获更为全面的上下文信息。实验结果表明,优化改进的YOLOv4算法mAP值达到了98.55%,比YOLOv4算法提高了1.9%,检测速度由35.66fps/s提升至53.25fps/s。确保准确度的前提,优化改进的YOLOv4算法在头盔配戴检测任务上的表现有了显著提升,使其更适用于轻量级应用场景。

In order to solve the problems of high missed detection rate and insufficient positioning accuracy of the current algorithm for helmet wearing detection of electric vehicle and motorcycle drivers in the traffic environment,an optimized and improved YOLOv4 algorithm is proposed.The introduction of MobileNetV2,a lightweight network,to replace the CSPDarknet53 backbone network,not only significantly reduces the number of parameters and computational costs,but also maintains the necessary characterization capabilities.The CA attention mechanism combined with the deep separable convolution is used to improve the representativeness of the network features,and the computational efficiency is further optimized by combining with the deep separable convolution.In order to make more efficient use of features at different levels and improve detection performance,the SPP+maximum pooling structure was introduced.Capture and fuse features at different spatial scales to facilitate model capture of more comprehensive contextual information.Experimental results show that the mAP value of the optimized YOLOv4 algorithm reaches 98.55%,which is 1.9%higher than that of the YOLOv4 algorithm,and the detection speed is increased from 35.66fps/s to 53.25fps/s.On the premise of ensuring accuracy,the optimized and improved YOLOv4 algorithm has significantly improved the performance of helmet wearing detection tasks,making it more suitable for lightweight application scenarios.