摘要
在公共安防领域,汽车底部潜藏的危险品危害性强,检测难度大.当前车底危险品检测主要通过模板匹配等传统目标检测技术进行检测,但存在检测速度慢、检测精度低的问题,为了能够更好地检测出藏匿于车底部位的危险品目标,提出一种改进的YOLOv3目标检测算法.该方法分别从多尺度图像训练、增加Inception-res模块和省去大尺度特征输出分支3个方面对YOLOv3网络进行改进.实验证明:在自制危险品数据集下,采用双数据集多尺度图像训练,网络的m AP值大约提高了0.9%,单张图像检测耗时大致不变;在3个支路分别增加相应Inception-res结构,网络的m AP值大约提高了1.5%,但是单张图像检测耗时却增加了原来的2.6倍;省去大尺度特征输出分支,网络的m AP值降低了0.3%,但是单张图像检测耗时也相应降低25.4%.通过结合上述方法对YOLOv3算法模型进行综合改进,选取双数据集多尺度图像训练的方式,同时省去大尺度特征输出分支,并在其他两支路增加相应Inception-res结构.这样在充分结合Inception-res结构优势的情况下,省去对检测耗时影响较大且对检测结果 m AP值影响较小的大尺度特征输出分支.实验测得改进网络m AP值大约提高2.2%左右,而单张图像检测耗时增加了0.014 s,在可接受范围内.且网络对于小尺寸目标识别效果明显增强,很好地满足了车底复杂背景危险品检测要求.
In the field of public security,dangerous objects hidden at the bottom of a vehicle are highly harmful and difficult to detect. In the field of vehicle bottom dangerous object detection,traditional object detection technology,such as template matching,is mainly used.However,the detection speed is slow and the detection accuracy is low.To better detect dangerous objects hidden at the bottom of a vehicle,an improved YOLOv3 detection algorithm is proposed. The method improves three aspects of the YOLOv3 network,i.e.,multi-scale image training,adding the Inception-res module,and eliminating the large-scale feature output branch. The experiment proves that,under the self-made dangerous object dataset,using the double-dataset multi-scale image training,the mAP value of the network increases by approximately 0.9%,but the detection time of a single image remains roughly the same. When adding the corresponding Inception-res structure to the three branches,the network’s mAP value increases by approximately 1.5%,but the detection time of a single image increases by 2.6 times. When eliminating the large-scale feature output branch,the network’s mAP value decreases by 0.3%,but the detection time of a single image decreases by 25.4%. By combining the three aspects,i.e.,adopting the double-dataset multi-scale image training method,eliminating the large-scale feature output branch,and adding the corresponding Inception-res structure to the two other branches,the YOLOv3 algorithm model is comprehensively improved. In this manner,in combination with the advantages of the Inception-res structure,the large-scale feature output branch that has a considerable effect on the detection time and has only a slight influence on the mAP value of the detection result is omitted. The experimental results show that the mAP value of the improved network increases by approximately 2.2%. Meanwhile,the detection time of a single image increases by 0.014 s,which is within the acceptable range. Moreover,the network has significantly enhanced the
作者
高春艳
赵文辉
张明路
孟宪春
Gao Chunyan;Zhao Wenhui;Zhang Minglu;Meng Xianchun(School of Mechanical Engineering,Hebei University of Technology,Tianjin 300130,China)
出处
《天津大学学报（自然科学与工程技术版）》
EI
CSCD
北大核心
2020年第4期358-365,共8页
Journal of Tianjin University：Science and Technology
基金
国家重点研发计划资助项目(2017YFC0806503).
关键词
深度学习
卷积神经网络
YOLOv3算法
危险品检测
deep learning
convolutional neural network
YOLOv3 algorithm
dangerous object detection
