基于深度学习的行人流量统计算法研究
发布时间：2018-09-13 07:42
【摘要】：近年来,计算机视觉技术逐渐成熟,其在智能监控领域的应用愈加广泛。大量原本需要人工完成的工作都可以由视觉算法来替代,极大地节约了人力成本。而在智能监控领域中,行人流量统计这一技术在商场、校园等场合都有着广泛的需求和应用,因此设计出一套智能的行人流量统计算法是十分有必要的。另一方面,如果能够将近期兴起的深度学习技术应用于其中,则将极大地提升算法的性能。本文研究并设计了一种基于深度学习的行人流量统计算法。本文通过综合应用基于深度学习的目标检测算法、单目标跟踪算法、数据关联算法等方法,设计了一套框架为“检测-跟踪-关联”的算法,用来完成对监控视频中行人流量的统计。本文主要进行了以下研究工作:首先,本文探究了行人流量统计的应用背景并阐述了研究的意义,然后分析了行人流量统计技术和基于深度学习的目标检测算法的发展现状,接下来阐述了本文的主要研究内容和研究方案,在研究方案中给出设计好的总体算法和框架。在制定了较为完善的研究方案的基础上,本文首先研究了卷积神经网络的组成结构和优化方法。然后回顾了传统目标检测算法和近年来产生的基于卷积神经网络的目标检测算法,最后确定使用SSD算法作为行人流量统计中的目标检测算法。接下来,本文研究了SSD算法的框架与原理,包括网络结构、缺省框的选择和训练目标函数等。之后,重点研究了SSD中的基网络,参照SSD原始基网络VGG和流行的CNN网络结构ZF-Net和SqueezeNet,重新设计了两种基网络并与VGG进行比较,结合实际需求最终确定了还是使用VGG作为基网络。在完成了对基于卷积神经网络的检测算法的研究后,本文还研究了需要使用的跟踪算法、数据关联算法和轨迹分析算法。确定了使用KCF算法作为跟踪算法,并直接使用OpenCV中的跟踪库。关联算法选取简单快速的基于距离的关联算法。最后设计了轨迹分析算法来实现双方向的计数。完成上述工作后,算法便已经完整。最后,阐述了实际的操作,包括摄像头的架设和样本视频的采集、检测图像数据集的制作、SSD检测器的训练、跟踪算法的实现、关联与轨迹分析算法的设计要点。然后使用设计的算法对所有样本视频进行分析,采用一些性能指标对其表现进行评价,其中平均识别率达到了96.24%,平均误检率为2.19%,平均漏检率为3.76%,全部视频平均帧率为24.09。结果表明所设计的算法可以能够满足项目的需求。
[Abstract]:In recent years, computer vision technology is gradually mature, its application in the field of intelligent monitoring is becoming more and more extensive. A large amount of manual work can be replaced by visual algorithm, which greatly saves manpower cost. In the field of intelligent monitoring, pedestrian flow statistics technology has a wide range of needs and applications in shopping malls, campus and other occasions, so it is necessary to design a set of intelligent pedestrian flow statistics algorithm. On the other hand, if the recently developed depth learning technology can be applied to it, the performance of the algorithm will be greatly improved. In this paper, a pedestrian flow statistic algorithm based on depth learning is studied and designed. In this paper, a set of algorithms called "detection, tracking and association" is designed by synthesizing the methods of target detection based on depth learning, single target tracking and data association. Used to complete the monitoring video traffic statistics. The main work of this paper is as follows: first, this paper explores the application background of pedestrian flow statistics and expounds the significance of the research, and then analyzes the development status of pedestrian flow statistics technology and target detection algorithm based on depth learning. Then, the main research content and research scheme of this paper are described, and the overall algorithm and framework are given in the research scheme. On the basis of a more perfect research scheme, this paper first studies the composition structure and optimization method of convolution neural network. Then the traditional target detection algorithm and the target detection algorithm based on convolution neural network are reviewed. Finally, the SSD algorithm is used as the target detection algorithm in pedestrian flow statistics. Then, this paper studies the framework and principle of SSD algorithm, including network structure, selection of default frame and training objective function. After that, the base network in SSD is studied emphatically. Referring to the original SSD network VGG and the popular CNN network structure ZF-Net and SqueezeNet, two base networks are redesigned and compared with VGG. Finally, VGG is used as the base network according to the actual requirements. After completing the research on the detection algorithm based on convolution neural network, this paper also studies the tracking algorithm, data association algorithm and trajectory analysis algorithm that need to be used. The KCF algorithm is used as the tracking algorithm, and the trace library in OpenCV is used directly. The association algorithm selects the simple and fast distance based association algorithm. Finally, a trajectory analysis algorithm is designed to realize double direction counting. After the above work has been completed, the algorithm is complete. Finally, the practical operation, including the installation of camera and the collection of sample video, the training of SSD detector, the realization of tracking algorithm and the design of correlation and trajectory analysis algorithm are discussed. Then the designed algorithm is used to analyze all the sample video, and some performance indexes are used to evaluate the performance of the video. The average recognition rate is 96.24, the average false detection rate is 2.19, the average missed detection rate is 3.76, and the average frame rate of the whole video is 24.09. The results show that the algorithm can meet the requirements of the project.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
[Abstract]:In recent years, computer vision technology is gradually mature, its application in the field of intelligent monitoring is becoming more and more extensive. A large amount of manual work can be replaced by visual algorithm, which greatly saves manpower cost. In the field of intelligent monitoring, pedestrian flow statistics technology has a wide range of needs and applications in shopping malls, campus and other occasions, so it is necessary to design a set of intelligent pedestrian flow statistics algorithm. On the other hand, if the recently developed depth learning technology can be applied to it, the performance of the algorithm will be greatly improved. In this paper, a pedestrian flow statistic algorithm based on depth learning is studied and designed. In this paper, a set of algorithms called "detection, tracking and association" is designed by synthesizing the methods of target detection based on depth learning, single target tracking and data association. Used to complete the monitoring video traffic statistics. The main work of this paper is as follows: first, this paper explores the application background of pedestrian flow statistics and expounds the significance of the research, and then analyzes the development status of pedestrian flow statistics technology and target detection algorithm based on depth learning. Then, the main research content and research scheme of this paper are described, and the overall algorithm and framework are given in the research scheme. On the basis of a more perfect research scheme, this paper first studies the composition structure and optimization method of convolution neural network. Then the traditional target detection algorithm and the target detection algorithm based on convolution neural network are reviewed. Finally, the SSD algorithm is used as the target detection algorithm in pedestrian flow statistics. Then, this paper studies the framework and principle of SSD algorithm, including network structure, selection of default frame and training objective function. After that, the base network in SSD is studied emphatically. Referring to the original SSD network VGG and the popular CNN network structure ZF-Net and SqueezeNet, two base networks are redesigned and compared with VGG. Finally, VGG is used as the base network according to the actual requirements. After completing the research on the detection algorithm based on convolution neural network, this paper also studies the tracking algorithm, data association algorithm and trajectory analysis algorithm that need to be used. The KCF algorithm is used as the tracking algorithm, and the trace library in OpenCV is used directly. The association algorithm selects the simple and fast distance based association algorithm. Finally, a trajectory analysis algorithm is designed to realize double direction counting. After the above work has been completed, the algorithm is complete. Finally, the practical operation, including the installation of camera and the collection of sample video, the training of SSD detector, the realization of tracking algorithm and the design of correlation and trajectory analysis algorithm are discussed. Then the designed algorithm is used to analyze all the sample video, and some performance indexes are used to evaluate the performance of the video. The average recognition rate is 96.24, the average false detection rate is 2.19, the average missed detection rate is 3.76, and the average frame rate of the whole video is 24.09. The results show that the algorithm can meet the requirements of the project.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前10条
1 韩娜;陈东伟;钟卓成;;智能视频客流人数统计系统的算法比较研究[J];信息技术;2016年06期
2 肖江;丁亮;束鑫;张文章;;一种基于计算机视觉的行人流量统计方法[J];信息技术;2015年08期
3 徐超;高梦珠;查宇锋;曹利民;;基于HOG和SVM的公交乘客人流量统计算法[J];仪器仪表学报;2015年02期
4 潘泓;李晓兵;金立左;夏良正;;一种基于二值粒子群优化和支持向量机的目标检测算法[J];电子与信息学报;2011年01期
5 李娟;杨录;张艳花;;基于红外传感器的楼宇人数统计系统的设计[J];山西电子技术;2010年06期
6 叶林;陈岳林;林景亮;;基于HOG的行人快速检测[J];计算机工程;2010年22期
7 田牛;应捷;;基于Canny边缘帧差法的公交客流量统计[J];微计算机信息;2010年32期
8 周晨卉;王生进;丁晓青;;基于局部特征级联分类器和模板匹配的行人检测[J];中国图象图形学报;2010年05期
9 余莉;韩方剑;;基于分水岭变换和遗传算法的自动目标检测[J];中国图象图形学报;2008年09期
10 李同治;丁晓青;王生进;;利用级联SVM的人体检测方法[J];中国图象图形学报;2008年03期
相关硕士学位论文 前10条
1 李娟;基于KCF的视频中运动物体的跟踪系统[D];湖南师范大学;2016年
2 蔡泽彬;基于视频分析的行人检测及统计方法研究[D];华南理工大学;2015年
3 g罘，
本文编号：2240506
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2240506.html
