【作者】 冯柯；

【导师】 陈临强；

【作者基本信息】 杭州电子科技大学 ， 计算机应用技术， 2009， 硕士

【摘要】 随着网络、通信和微电子技术的飞速发展,一些特定功能的视觉分析系统以其直观、方便和内容丰富等特点,日益受到人们的亲睐,其中交通监控领域应用最为广泛。然而,全天候监控捕获的大量视频信息,若采用人工搜索方法来寻找目标,不仅低效,由人为因素造成的失误亦是难以避免。因此,人们希望计算机能具有类似人类视觉系统的能力,可以分析、理解图像或视频的内容,以实现视频分析系统的智能化、实用化,视频分析技术应运而生。视频分析技术主要是处理包含各种运动目标的视频序列,从场景中检测、跟踪、分类识别目标,并对其行为进行理解和描述。其中,目标分类是基于视频的运动分析课题中的一个重要方面,其研究内容是在运动检测和跟踪的基础上,依据提取的运动目标区域形状特征和运动属性,对运动目标区域进行语义上的分类。目标分类技术研究对更高层次的视频理解技术的发展有重要意义。目前国外研究机构和国内高校在目标分类技术上取得了一定进展,但仍存在一些应用上的限制和不足,其中目标分类的视点依赖性问题是影响分类稳定性的主要因素。所谓视点依赖性是指目标的2D特征在投影到图像平面时发生了透视形变,从而导致其无法准确的用于分类。本文围绕运动目标分类及其在交通视频目标检索中的应用这一课题,重点阐述分类过程中视点依赖性问题的解决方法,并对各类相关技术进行了研究,具有重要的理论意义和实际价值。本文的工作主要分以下几个部分:(1)介绍目标分类领域基础理论及相关研究,包括目标的特征表达、目标分类方法、运动目标检测与跟踪以及场景知识在目标分类中的作用。(2)提出了一种基于kalman预估模型和最大化后验概率匹配的粘连目标跟踪方法,实现了目标相互遮挡时连续稳定跟踪。(3)针对目标分类中遇到的视点依赖性问题,介绍当前具有代表性的三种目标2D特征透视变形的矫正方法,并提出了基于地平面矫正的目标2D特征恢复算法。在标准化后特征的基础上,采用多类支持向量机实现视点无关运动目标分类。(4)研究基于目标特征和语义类别的运动目标检索方法,对目标特征数据组织形式、目标检索方式、查询结果显示方式等进行了探讨;(5)本文搭建了基于Visual C++平台和OpenCV图像处理库的实验环境,利用程序验证了提出了的各类算法,实验证明本文算法实现了不同视点下运动目标的准确分类。更多还原

【Abstract】 As the rapid development of the network, communications,microelectronic technology,a number of visual analysis system with specific function get the growing popularity of the pro-gaze for its intuitive, convenient and rich in content. Traffic monitoring is the most widely used area. However, the all-weather surveillance system captured a large number of video information, in which it is inefficient and mistakable to find the target with the use of pure-manual search methods. Therefore, it was hoped that the computer can analyze and understand video content, in order to achieve video analysis intelligent and practical system, video analysis techniques have emerged.Video analysis techniques deal primarily with video sequences which contain a variety of moving target, and detect, track, and classify and recognition goals, then understand and describe their behavior. Among them, target classification is an important aspect of the Video analysis with the content of classifying the object area based on motion analysis by their features of shape and motion, and is important to the development of high level video understanding techniques.Currently overseas research institutions and national universities have made some progress in the target classification technology, but there are still some constraints and lack of applications, in which the view-dependent issue of classification is the main factor affecting the stability of classification. The so-called view-dependent issue means 2D feature of object has some distortion because of projection, causing it is not correct for classification. The major work is described as follows:(1) Introduce basic theory of target classification and related research, including the expression of object features, common methods of object classification, moving target detection and tracking, and the use of scene knowledge.(2) The predict model based on kalman filter combine with maximum posterior probability for object matching is proposed to track moving objects with occlusions.(3) For the view-dependent problems encountered in target classification, Describes three representative kinds of 2D features rectification methods,and then promote 2D features recovery algorithm based on ground plane rectification. Based on the normalized characteristics, using Multi-Class Support Vector Machines (SVM) to achieve view-independent classification of moving object.(4) Research on moving object retrieval method based on video features,discuss the organization of feature data,the way of object retrieval and display form of retrival results.(5) By building the experiment platform based on Visual C++ and OpenCV graphics library. We verify the various algorithms with coding. Results shows the algorithm solve the view-dependent problems in object classification and achieve view-independent object classification.更多还原