基于改进YOLOv5车辆检测方法
Vehicle detection method based on improved YOLOv5
吕宏泽 1李继财 2杨乔楠 1陈学永 1李西兵1
作者信息
- 1. 福建农林大学机电工程学院,福建福州 350100;福建农林大学福建省农业信息感知技术重点实验室,福建福州 350100
- 2. 山东劳动职业技术学院智能制造系,山东济南 250022
- 折叠
摘要
针对现有目标检测在智能交通系统和自动驾驶等领域存在车辆目标检测精度低、鲁棒性较差等问题,提出一种基于YOLOv5的车辆目标检测算法.在YOLOv5s网络模型框架中,添加注意力机制增强特征,提取重要特征;添加小目标检测层提升对遮挡重叠弱小目标识别的准确率;引入金字塔池化(SPPFCSPC),提高网络空间特征提取能力;引入损失函数(SIoU_Loss)加快边界框回归速率,提高定位精度,消除重叠检测.基于自制车辆检测数据集进行实验,其结果表明,改进网络模型与原YOLOv5s网络模型相比,不同目标类的平均准确率均有明显提高,平均准确率均值提升3.25％,查准率提高4.14％,召回率提高3.05％,检测速度满足实时性要求.
Abstract
In areas such as intelligent transport systems and autonomous driving,existing target detection suffers from low accu-racy and poor robustness of vehicle target detection.A YOLOv5 based algorithm for vehicle target detection was proposed.In the YOLOv5s network model framework,attention mechanisms was added to enhance features and extract important features.A small target detection layer was added to improve the accuracy of recognition of occluding overlapping weak targets.Pyramid pooling(SPPFCSPC)was introduced to improve network spatial feature extraction.A loss function(SIoU_Loss)was introduced to speed up the rate of bounding box regression,improve localization accuracy and eliminate overlap detection.Experiments were carried out based on a home-made vehicle inspection dataset.Compared with the original YOLOv5s network model,the results show that the average accuracy is improved by 3.25％,the average accuracy of different target classes is significantly improved,and the average accuracy is increased by 3.25％,the accuracy rate is increased by 4.14％,and the recall rate is increased by 3.05％.The detection speed meets the real-time requirements.
关键词
车辆检测/深度学习/损失函数/特征增强/图像处理/神经网络/智能交通Key words
vehicle detection/deep learning/loss function/feature enhancement/image processing/intelligent transportation/neural networks引用本文复制引用
基金项目
福建农林大学科技创新专项(CXZX2020132B)
福建省自然科学基金(2022J01609)
出版年
2024
NETL
NSTL
万方数据