摘要
车辆信息检测是车型识别在智慧交通领域中的首要任务。针对现有的车辆信息检测技术在检测速度、精度以及稳定性方面存在的问题,提出了基于YOLOv3的深度学习目标检测算法——YOLOv3-fass。该算法以DarkNet-53网络结构为基础,删减了部分残差结构,降低了卷积层的通道数,添加了1条下采样支路和3个尺度跳连结构,增加了一个检测尺度,并通过K-均值聚类与手动调节相结合的方法计算出12组锚框值。最后通过迁移学习机制对YOLOv3-fass算法进行微调。在自研的车辆数据集上,YOLOv3-fass算法与YOLOv3、YOLOv3-tiny、YOLOv3-spp算法以及具有ResNet50和DenseNet201经典网络结构的算法做了对比实验,结果表明YOLOv3-fass算法能够更精准、高效、稳定地检测到车辆信息。
Vehicle information detection is the primary task of vehicle type identification in the field of intelligent transportation.Based on deep learning YOLOv3(You Only Look Once Version 3)model,a new YOLOv3-fass object detection algorithm was proposed to address some problems existing in vehicle information detection technology such as detection speed,accuracy and stability.In this improved algorithm,based on DarkNet-53 network structure,some residual structures were deleted,and a number of channels of convolutional layer were reduced;a down-sampling branch,three scale-hopping connection structures,and one detection scale were added;and twelve groups of anchor frame values were calculated through the means of K-means clustering algorithm combined with manual setting.Finally,YOLOv3-fass algorithm was fine-tuned through the migration learning mechanism of multi-stage pre-training.YOLOv3-fass algorithm was compared with YOLOv3,YOLOv3-tiny,YOLOv3-spp and two algorithms with ResNet50 and DenseNet201 on the vehicle data set.The experimental results show that YOLOv3-fass algorithm can detect vehicle information more accurately,efficiently and stably.
作者
冯加明
储茂祥
杨永辉
巩荣芬
FENG Jiaming;CHU Maoxiang;YANG Yonghui;GONG Rongfen(School of Electronics and Information Engineering,University of Science and Technology Liaoning,Anshan 114051,Liaoning,P.R.China)
出处
《重庆大学学报》
CSCD
北大核心
2021年第12期71-79,共9页
Journal of Chongqing University
基金
国家自然科学基金资助项目(71771112)
辽宁省自然科学基金资助项目(20180550067)。