摘要
基于深度学习的目标检测算法在智能交通的应用中,对于车辆检测存在模型参数量大、计算速度慢和简单网络精准度较低的问题。本文提出了一种高效的轻量化车辆检测模型,该检测模型采用YOLOv4网络作为参考模型进行改进。首先,本文采用CSPMobileViT网络来替换原始主干网络,然后将PANet替换成BiFPN,并且将BiFPN中的3×3标准卷积替换成深度可分离卷积,最后,在BiFPN之前和YOLO-Head之前添加ECA模块。在损失函数部分,将边框回归损失CIoU改进为Focal EIoU来解决难易样本不平衡的问题。实验结果表明改进网络的mAP值为96.77%,检测速度达到每张图片0.0234 s,模型大小只有32.76 MB,参数量为8587541,与原始算法相比mAP提升了1.54%,而模型大小和参数量仅约为原始模型1/8,并且FPS提升了7.5,改进算法具有更好检测效果。
In the application of Intelligent Transportation,target detection algorithm based on deep learning has the problems of large number of model parameters,slow calculation speed and low accuracy of simple network for vehicle detection.This paper presents an efficient lightweight vehicle detection model,which is improved by using YOLOv4 network as a reference model.First,this paper uses CSPMobileViT network to replace the original backbone network,then replaces PANet with BiFPN,replaces 3×3 standard convolution in BiFPN with deep detachable convolution,and finally adds ECA module before BiFPN and YOLO-Head.In the loss function section,the Border Regression Loss CIoU is improved to Focal EIoU to solve the problem of difficult sample imbalance.The experimental results show that the mAP value of the improved network is 96.77%,the detection speed reaches 0.0234 s per picture,the model size is only 32.76 MB,and the parameter amount is 8587541.Compared with the original algorithm,the mAP is improved by 1.54%,while the model size and number of parameters are only about 1/8 of the original model,and the FPS is improved by 7.5,so the improved algorithm has better detection effect.
作者
郑玉珩
黄德启
Zheng Yuheng;Huang Deqi(School of Electrical Engineering,Xinjiang University,Urumqi 830017,China)
出处
《电子测量技术》
北大核心
2023年第2期175-183,共9页
Electronic Measurement Technology
基金
国家自然科学基金(51468062)项目资助
作者简介
黄德启,副教授,硕士生导师,主要研究方向为智能交通。E-mail:dqhuang88@qq.com;郑玉珩,硕士研究生,主要研究方向为深度学习、图像处理、智能交通。E-mail:1812294492@qq.com
