摘要
为了对汽车试验场的外物入侵进行识别并预警,结合现有监控系统,利用Tensorflow搭建Faster-RCNN框架,分别使用Inception V2和Res Net101作为核心卷积神经网络,用自建的标注数据集进行最终训练,利用得到的模型进行迁移测试,试验结果表明:Inception V2系统的平均精度值为81.7%,Res Net101系统的平均精度值为84.1%。将两种系统结合Opencv的图像抓取功能及现有摄像监控设备进行联合测试,结果表明两种系统均能在阴天和低像素摄像头搭配下实时对高速或慢速移动的物体进行识别、分类、标注、预警。
In order to recognize and alarm invaders which entered into automobile proving ground,Faster-RCNN framework was established by Tensorflow,combining with current video monitoring system. Inception V2 and Res Net101 were taken as core convolution neural network respectively. The final training was carried out with the self-built annotation data set,and the migration test was carried out with the obtained model. The experimental results show that the average precision value of Inception V2 system and Res Net101 system reaches 81.7% and 84.1% respectively. The joint test combining the two systems with the image capture function of Opencv and current video monitoring system was carried out. The results prove that both systems can recognize,classify,mark and alarm the high-speed or low-speed moving objects in real time under cloudy weather or low-pixel camera condition.
作者
向华荣
曾敬
XIANG Huarong;ZENG Jing(Chongqing Xibu Automotive Proving Ground Management Co.,Ltd.,Chongqing 408300,China;China Automotive Engineering Research Institute Co.,Ltd.,Chongqing 401122,China)
出处
《重庆交通大学学报（自然科学版）》
CAS
CSCD
北大核心
2020年第1期8-14,共7页
Journal of Chongqing Jiaotong University(Natural Science)
基金
工业强基工程项目(0714-EMTC02-5593/20)
关键词
车辆工程
汽车试验场
外物入侵
卷积神经网络
目标检测
vehicle engineering
automobile proving ground
invaders
convolution neural network
target detection
