节点文献

基于改进Mask R-CNN的交通监控视频车辆检测算法

Vehicle Detection Algorithm Based on Improved Mask R-CNN in Traffic Surveillance Video

  • 推荐 CAJ下载
  • PDF下载
  • 不支持迅雷等下载工具,请取消加速工具后下载。

【作者】 陆飞沈世斌苏晓云谢非章悦刘益剑

【Author】 Lu Fei;Shen Shibin;Su Xiaoyun;Xie Fei;Zhang Yue;Liu Yijian;School of Electrical and Automation Engineering,Nanjing Normal University;Jiangsu Key Laboratory of 3D Printing Equipment and Manufacturing;Nanjing Industry Institute for Advanced Intelligent Equipment Co.,Ltd.;

【通讯作者】 沈世斌;

【机构】 南京师范大学电气与自动化工程学院江苏省三维打印装备与制造重点实验室南京智能高端装备产业研究院有限公司

【摘要】 针对交通监控视频车辆检测常易受到遮挡导致目标车辆出现漏检或误检的问题,提出一种基于改进Mask R-CNN的交通监控视频车辆检测算法.采用基于bottleneck结构的主干网络,提高主干网络提取特征的能力;通过基于预测mask分数的掩码分支,融合目标的类别分数和掩码质量分数,提高车辆的掩码质量;通过基于Arcface Loss的目标检测损失函数设计,提高不同特征之间的可判别性,提高目标的检测精度.实验结果表明,改进的Mask R-CNN模型可更好地检测到被遮挡的车辆,目标车辆的检测精度超过Faster R-CNN、YOLO v3和Mask R-CNN模型,可解决目标车辆漏检或误检问题.

【Abstract】 Aiming at the problem of missing detection or wrong detection of target vehicles caused by occlusion in traffic surveillance video,an improved vehicle detection algorithm based on Mask R-CNN traffic surveillance video is proposed.Firstly,the backbone network based on the bottleneck structure is used to improve the ability of extracting features from the backbone network. Then,the mask branch based on the predicted mask score is used to fuse the target’s category score and mask quality score to improve the vehicle’s mask quality. Finally,the target detection loss function based on Arcface Loss can improve the discriminability between different features and improve the detection accuracy of the target.The experimental results show that the improved Mask R-CNN model can better detect the shielded vehicle,and that the detection accuracy of the target vehicle is higher than those of the Faster R-CNN,YOLO v3 and Mask R-CNN model,thus solving the problem of missing or wrong detection of the target vehicle.

【基金】 国家自然科学基金项目(61601228、41974033、61803208);江苏省自然科学基金项目(BK20161021、BK20180726);江苏省高校自然科学基金项目(17KJB510031)
  • 【文献出处】 南京师范大学学报(工程技术版) ,Journal of Nanjing Normal University(Engineering and Technology Edition) , 编辑部邮箱 ,2020年04期
  • 【分类号】TP183;TP391.41;U495
  • 【被引频次】4
  • 【下载频次】227
节点文献中: 

本文链接的文献网络图示:

本文的引文网络