低分辨率交通视频中运动物体识别算法研究
发布时间：2018-12-31 19:26
【摘要】：随着计算机技术、视频技术和图像处理技术的不断发展,基于图像处理方法的运动物体识别技术被越来越多地运用于智能视频监控系统之中。相对于传统的识别方法,图像处理方法因为其不需要增设外部设备,系统整体成本较低,可扩展性强等诸多优点,正越来越多的受到研究人员和实际工程人员的关注。交通视频监控是智能视频监控的一个重要应用领域。 交通视频自身具有很多特点,如视频分辨率较低、运动物体背景复杂、气象条件多变、存在大范围光线变化等。本文对当前存在的运动物体检测、跟踪和分类方法,特别是基于视频图像处理的方法进行了调研,分析了各种方法的适用场景和优缺点,并针对本文实际的应用场景——低分辨率交通视频图像——提出了一种基于多特征融合和多帧融合的运动物体识别算法。 本文首先对运动物体的分割方法进行研究,通过对交通视频建立合适的背景模型,采用背景差分的方法提取出运动物体。利用形态学处理等方法去除背景噪声的干扰,并提出了一种基于区域生长的阴影去除方法,以获得较为准确的运动物体。接着,提取出运动物体的几何特征和运动特征,再分别基于支持向量机和级联分类器两种策略对运动物体的特征进行融合,获取运动物体的单帧判决信息。最后融合视频序列中多帧的判决信息,完成对运动物体的识别,得到最终分类信息。本文给出的这种基于多特征融合和多帧融合的运动物体识别算法,经实验证明,在实际低分辨率交通视频的应用场景中,可以较好地识别出运动物体,且计算复杂度低,能够满足交通视频的实时处理需求。 在完成算法设计后,本文对智能交通视频监控系统的构建进行了描述,简要论述了该系统的组成原理、处理流程和实现方案。
[Abstract]:With the development of computer technology, video technology and image processing technology, moving object recognition technology based on image processing method is more and more used in intelligent video surveillance system. Compared with the traditional recognition methods, the image processing method is paid more and more attention by researchers and practical engineers because it does not need to add external equipment, the overall cost of the system is relatively low, the expansibility is strong and so on. Traffic video surveillance is an important application field of intelligent video surveillance. Traffic video has many characteristics, such as low video resolution, complex background of moving objects, changeable meteorological conditions, large range of light changes and so on. In this paper, the existing methods of moving object detection, tracking and classification, especially based on video image processing, are investigated, and the applicable scenes, advantages and disadvantages of these methods are analyzed. And a moving object recognition algorithm based on multi-feature fusion and multi-frame fusion is proposed for the actual application scene of this paper, which is low resolution traffic video image. In this paper, the segmentation method of moving object is studied firstly, and the moving object is extracted by using background difference method by establishing a suitable background model for traffic video. Morphological processing is used to remove background noise, and a shadow removal method based on region growth is proposed to obtain more accurate moving objects. Then, the geometric features and motion features of moving objects are extracted, and then the features of moving objects are fused based on support vector machine and cascade classifier, respectively, and the single frame decision information of moving objects is obtained. Finally, the decision information of multiple frames in the video sequence is fused to complete the recognition of moving objects and the final classification information is obtained. The motion object recognition algorithm based on multi-feature fusion and multi-frame fusion presented in this paper has been proved by experiments that it can recognize moving objects well in the application scene of actual low-resolution traffic video, and the computational complexity is low. It can meet the need of real-time traffic video processing. After completing the algorithm design, this paper describes the construction of intelligent transportation video surveillance system, and briefly discusses the principle of the system, processing process and implementation scheme.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN948.6;TP391.41
本文编号：2397015
[Abstract]:With the development of computer technology, video technology and image processing technology, moving object recognition technology based on image processing method is more and more used in intelligent video surveillance system. Compared with the traditional recognition methods, the image processing method is paid more and more attention by researchers and practical engineers because it does not need to add external equipment, the overall cost of the system is relatively low, the expansibility is strong and so on. Traffic video surveillance is an important application field of intelligent video surveillance. Traffic video has many characteristics, such as low video resolution, complex background of moving objects, changeable meteorological conditions, large range of light changes and so on. In this paper, the existing methods of moving object detection, tracking and classification, especially based on video image processing, are investigated, and the applicable scenes, advantages and disadvantages of these methods are analyzed. And a moving object recognition algorithm based on multi-feature fusion and multi-frame fusion is proposed for the actual application scene of this paper, which is low resolution traffic video image. In this paper, the segmentation method of moving object is studied firstly, and the moving object is extracted by using background difference method by establishing a suitable background model for traffic video. Morphological processing is used to remove background noise, and a shadow removal method based on region growth is proposed to obtain more accurate moving objects. Then, the geometric features and motion features of moving objects are extracted, and then the features of moving objects are fused based on support vector machine and cascade classifier, respectively, and the single frame decision information of moving objects is obtained. Finally, the decision information of multiple frames in the video sequence is fused to complete the recognition of moving objects and the final classification information is obtained. The motion object recognition algorithm based on multi-feature fusion and multi-frame fusion presented in this paper has been proved by experiments that it can recognize moving objects well in the application scene of actual low-resolution traffic video, and the computational complexity is low. It can meet the need of real-time traffic video processing. After completing the algorithm design, this paper describes the construction of intelligent transportation video surveillance system, and briefly discusses the principle of the system, processing process and implementation scheme.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN948.6;TP391.41
【参考文献】
相关期刊论文 前10条
1 闻帆;屈桢深;闫纪红;;集成多特征信息的运动阴影检测[J];哈尔滨工业大学学报;2011年05期
2 魏伟波;芮筱亭;;图像边缘检测方法研究[J];计算机工程与应用;2006年30期
3 陈柏生;陈锻生;;基于归一化rgb彩色模型的运动阴影检测[J];计算机应用;2006年08期
4 盛能;王慧;刘泓;;混合交通流中的自行车识别及参数提取[J];计算机应用研究;2010年05期
5 赵来刚;陈道炯;;一种基于Sobel算子的新型边缘提取技术[J];机械与电子;2011年02期
6 张嘉;梁茵;;论我国城市道路建设与规划管理问题[J];四川建筑;2006年04期
7 余孟泽;刘正熙;骆键;杨丹;;融合纹理特征和阴影属性的阴影检测方法[J];计算机工程与设计;2011年10期
8 李子龙;刘伟铭;;基于组合特征和HSI颜色空间的阴影检测[J];微电子学与计算机;2011年03期
9 高岚;董慧颖;兰利宝;;自适应背景下运动目标阴影检测算法研究[J];现代电子技术;2008年06期
10 ;中国安防行业“十二五”(2011～2015年)发展规划[J];中国安防;2011年03期
相关博士学位论文 前2条
1 李峰;智能视频监控系统中的行人运动分析研究[D];中国科学技术大学;2011年
2 陈功;鲁棒的智能视频监控方法研究[D];中国科学技术大学;2008年
，本文编号：2397015
本文链接：https://www.wllwen.com/kejilunwen/wltx/2397015.html
