摘要
实时检测道路中的车辆并统计其数量是道路监控系统的重要任务之一,道路上车流量较大时,车辆之间相互遮挡,难以对道路监控视频中的车辆做到实时准确的检测。为此,提出以车窗代替车体作为目标物进行检测,由此提高被遮挡车辆的检出率。利用残差连接和多尺度特征提取方法,构建了只有24个卷积层的检测网络。实验结果表明,该检测算法的召回率达到了95%,正确率达到了96%,在GeForce GTX 1080Ti显卡环境下的检测速度达到了142FPS。这种方法可有效提高对相互遮挡的车辆的检出率,且可对高清监控视频中车辆进行实时检测。
Real-time vehicle detection and counting is one of the significant tasks of road monitoring system.Accurate vehicle detection in real time surveillance video is tough in the large traffic flow because of the mutual occlusion.Window replaces body of the car is proposed as the solution that improves the detection rate of mutual occlusion vehicles effectively.A detection network with only 24 convolution layers construct by residual connection and multi-scale feature extraction.The experimental results illustrate the recall rate of the detection algorithm reaches 95%,the accuracy rate reaches 96% and the detection speed reaches 142 FPS in GeForce GTX 1080 Ti.The proposed arithmetic improves the detection rate of mutual occlusion vehicles effectively and has a strong ability in real-time detection of vehicles in high-definition surveillance video.
作者
王亮亮
王国栋
赵毅
潘振宽
WANG Liang-liang;WANG Guo-dong;ZHAO Yi;PAN Zhen-kuan(College of Computer Science and Technology, Qingdao University, Qingdao 266071,China)
出处
《青岛大学学报（自然科学版）》
CAS
2019年第3期1-7,共7页
Journal of Qingdao University(Natural Science Edition)
基金
国家自然科学基金面上项目(批准号:61772294)资助
"十二五"国家科技支撑计划(批准号:2014BAG03B05)资助
关键词
计算机视觉
卷积神经网络
残差连接
多尺度特征提取
创新数据标注
computer vision
convolution neural network
residual connection
multi-scale feature extrac-tio n
innovation data tagging
