基于改进的SECOND网络与MobileNet V2网络的视频监控与追踪技术实现  

Implementation of Video Monitoring and Tracking Technology Based on Improved SECOND Network and MobileNet V2 Network

作　　者：林自强 李明 刘张榕[2] LIN Ziqiang;LI Ming;LIU-ZHANG Rong(Zhongshan Power Supply Bureau,Guangdong Power Grid Co.,Ltd.,Zhongshan 528400,China;Department of Information Engineering,Fujian Forestry Vocational&Technical College,Nanping 353000,China)

机构地区：[1]广东电网有限责任公司中山供电局,广东中山528400 [2]福建林业职业技术学院信息工程系,福建南平353000

出　　处：《测控技术》2024年第5期78-84,共7页Measurement & Control Technology

基　　金：福建省中青年教师教育科研项目(科技类)(JAT220588)。

摘　　要：为提高作业现场视频监控水平,提出一种基于深度学习的视频监控与追踪方法。该方法通过采用Res2SENet网络作为SECOND网络的卷积模块,并使用可变形卷积替代SECOND网络的标准卷积,以实现SECOND网络的改进。首先,基于改进的SECOND网络对目标进行检测,实现作业现场视频监控目标检测;然后,通过采用空洞卷积替代MobileNet V2网络的普通卷积,并使用改进后的MobileNet V2网络作为目标追踪算法,实现作业现场视频监控目标检测与跟踪;最后,在典型的包含大量激光点云图像的KITTI数据集上进行测试。结果表明,该方法利用改进SECOND网络对作业现场视频监控三维目标检测的平均精度和检测时间分别为81.62%和0.048 s,相较于标准SECOND网络、特征金字塔网络(Feature Pyramid Network,FPN)、F-PointNet网络,改进SECOND网络具有明显优势;利用改进的Mo-bileNet V2网络对作业现场视频监控三维目标跟踪的准确度、精确度和跟踪数分别为81.62%、80.55%和57.30%,丢失数和跟踪轨迹中行人ID瞬间转换次数分别为11.08%和22%,具有较快的运行速度,为39 f/s,相较于MobileNet V2网络、马尔可夫决策过程(Markov Decision Process,MDP)网络、平滑支持向量机(Smooth Support Vector Machine,SSVM)网络,改进的MobileNet V2网络在各项指标上均具有一定优势,可以满足作业现场视频监控目标的检测与实时跟踪需求。In order to improve the job site video monitoring level,a video monitoring and tracking method based on deep learning is proposed.This method enables the improvement of the SECOND network by adopting Res2SENet network as the convolution module of SECOND network and using deformable convolution to re-place the standard convolution of SECOND network.Firstly,the target is detected based on the improved SEC-OND network,and the job site video monitoring target detection is realized.Then,the detection and tracking of job site video monitoring target is realized by adopting dilation convolution to replace the ordinary convolution·78·of the MobileNet V2 network and using the improved MobileNet V2 network as the target tracking algorithm.Finally,test is performed on a typical KITTI dataset containing a large number of laser point cloud images.The results show that the average accuracy and detection time of the method utilizing the improved SECOND net-work to detect the job site video monitoring 3D targets are 81.62%and 0.048 s,respectively,which means that the improved SECOND network has obvious advantages over the standard SECOND network,feature pyra-mid network(FPN)network and F-PointNet network.The accuracy,precision and tracking number using the improved MobileNet V2 network to detect the job site video monitoring 3D targets are 81.62%,80.55%and 57.30%,respectively.The number of lost and instantaneous conversion times of pedestrian ID in the tracking trajectory are 11.08%and 22%,respectively.It has a faster operating speed,which is 39 f/s.Therefore,com-pared with MobileNet V2 network,MDP network and SSVM network,the improved MobileNet V2 network has certain advantages in various indicators,and can meet the detection and real-time tracking requirements of the job site video monitoring target.

关 键 词：深度学习 视频监控 目标检测 目标跟踪 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]