[1]张震,李浩方,李孟洲,等.改进YOLOv3算法与人体信息数据融合的视频监控检测方法[J].郑州大学学报(工学版),2021,42(01):28-34.[doi:10.13705/j.issn.1671-6833.2021.01.005]
 Zhang Zhen,Li Haofang,Li Mengzhou,et al.A New Method of Human Information Detection in Video Surveillance[J].Journal of Zhengzhou University (Engineering Science),2021,42(01):28-34.[doi:10.13705/j.issn.1671-6833.2021.01.005]
点击复制

改进YOLOv3算法与人体信息数据融合的视频监控检测方法()
分享到:

《郑州大学学报(工学版)》[ISSN:1671-6833/CN:41-1339/T]

卷:
42卷
期数:
2021年01期
页码:
28-34
栏目:
出版日期:
2021-03-14

文章信息/Info

Title:
A New Method of Human Information Detection in Video Surveillance
作者:
张震李浩方李孟洲马军强
郑州大学电气工程学院;

Author(s):
Zhang Zhen; Li Haofang; Li Mengzhou; Ma Junqiang;
School of Electrical Engineering, Zhengzhou University;

关键词:
Keywords:
DOI:
10.13705/j.issn.1671-6833.2021.01.005
文献标志码:
A
摘要:
针对目前社区视频监控使用人脸相机仅采集出入口人脸数据,而缺失有数据价值的人体其它信息的问题. 本文提出一种将改进YOLOv3网络和调用人体信息识别模块相结合的人体信息检测方法.采用K-means++算法获取数据集的先验框,选用新的边界框回归损失函数GIoU提高检测精度,再进行多尺度训练得到人体检测网络模型,最后利用人体检测模型在检测到人体目标后调用人体信息识别模块对人体信息进行分析和保存.实验结果表明,该方法既能快速检测人体目标,还能准确获取人体目标的各种属性信息.其中人体检测模型在测试集上的mAP(Mean Average Precision)达为91.8%,识别速率为45帧/s
Abstract:
For the current community video surveillance, the use of a face camera only collects the entrance and exit face data, and the lack of other valuable human information. This paper proposes a human information detection method that combines improved YOLOv3 network and calling human information recognition module. This paper uses the K-means ++ algorithm to obtain the prior frame of the data set, selects the new bounding box regression loss function GIoU to improve the detection accuracy, and then performs multi-scale training to obtain the human detection network model. Finally, the human detection model is used to detect human targets The human body information recognition module is called to analyze and save human body information. The experimental results show that the method can not only detect human targets quickly, but also accurately obtain various attribute information of human targets. The mAP (Mean Average Precision) of the human detection model on the test set i s 91.8%, and the recognition rate is 45 frames / s
更新日期/Last Update: 2021-03-15