中文摘要:
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MobileNet 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。
英文摘要:
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 replace the standard convolution of SECOND network.Firstly,the target is detected based on the improved SECOND 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 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 network 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 pyramid 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,compared 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.
到此这篇pointnet++网络原理（pointnet网络结构）的文章就介绍到这了,更多相关内容请继续浏览下面的相关
推荐文章，希望大家都能在编程的领域有一番成就！
版权声明：
本文来自互联网用户投稿，该文观点仅代表作者本人，不代表本站立场。本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。
如若内容造成侵权、违法违规、事实不符，请将相关资料发送至xkadmin@xkablog.com进行投诉反馈，一经查实，立即处理！
转载请注明出处，原文链接：https://www.xkablog.com/cjjbc/16375.html
