【作者】 王夏黎；

【导师】 周明全；

【作者基本信息】 西北大学 ， 计算机软件与理论， 2004， 博士

【摘要】 交通视频信息在交通监控和交通管理中一直作为重要内容被采集和利用。为充分利用采集的信息,提高交通监控和管理的智能化水平,以视频图像处理、分析、理解为基础的视频监视技术越来越引起人们的重视。视频监视技术是图像处理与计算机视觉领域的一个研究热点,它与传统监视技术的区别在于其智能性。因此,在智能交通系统中开展交通视频监视技术的研究有十分重要的现实意义。本文主要围绕交通视频监视技术的关键问题完成了以下几项工作。 1.分析比较了四种运动目标检测方法的原理和特点,针对各自方法的不足,做出了相应改进。为解决背景差方法中背景图像易受噪声干扰的问题,给出了三种背景更新方法来减少干扰。为了解决背景频繁变化情况下的目标检测问题,给出了建立背景模型的方法。为解决帧间差分法不能完整检测到运动目标的问题,给出了一种改进的帧间差分方法。为了克服光流法运算量大,实时性差的问题,给出了一种与帧间差分法相结合的改进的快速光流法。 2.针对复杂变化背景下运动目标的检测问题,研究了帧间像素的颜色共生性(颜色相关性)与频繁变化背景和前景之间的关系。根据帧间像素颜色共生性对于频繁变化背景有更多意义这一情况,研究了贝叶斯准则和颜色共生概率在即时图像变化情况下它们之间的联系,给出了用贝叶斯准则判别前景与背景像素的数学公式。研究了频繁变化背景模型的更新方法。提出用固定背景区域的参数背景图象和变化背景目标的像素颜色共生统计表来共同维护背景的方法。给出了长期和短期更新像素颜色共生统计表的策略。基于上述研究,给出了一种基于颜色共生性的,用贝叶斯决策准则从复杂的含有不固定目标的视频图像中检测前景目标的方法。 3.提出了一种改进的基于区域的多目标跟踪方法。该方法根据运动图像序列中帧间运动的连续性原则,建立匹配函数,并将其用于运动目标区域特征的匹配检测;使用Kalman滤波器对目标区域下一步出现的位置进行预测,缩小匹配搜索范围,加快搜索匹配速度;建立跟踪控制表来记录目标区域最新的运动参数,以保证运动目标跟踪的连续性;采用目标质心间距和延迟搜索法解决由于重复匹配标记与运动暂停造成的目标丢失现象;利用设置“入界区”与“出界区”来正确判断新旧目标的出现与消失。从而实现对目标的正确跟踪。 4.提出了一种用支撑矢量机(SVM)对交通目标进行分类的方法。该方法的分类思想是为每一类目标分别建立一个SVM分类器,然后用对应的目标样本进行训练,最后用训练好的SVM分类器进行分类。实验表明:该方法实现了对行人、小型车、大型车三类交通目标的准更多还原

【Abstract】 Because traffic video information has been collected and used as an important part of traffic surveillance and management, great importance has been attached to video surveillance technology which is based on video image processing, analyzing, and understanding, to improve intelligence of traffic surveillance and management. As a subject of general interest in area of image processing and computer vision, video surveillance technology is different from traditional surveillance technology in that it is highly intelligent; therefore, research in this technology and its application in Intelligent Transportation System are of great practical significance. Focusing on the key technological problems of traffic video surveillance, the main research work of this paper is presented as follows:1. The principles of four methods of moving objects detection are given. By comparing and analyzing different methods and their characteristics, the drawbacks of these methods are pointed out and ways of improvement are offered. The drawback of Background Subtract Method is that the background is very sensitive to noise. So three new methods of background renewal are given to reduce the disturbance of noise. To solve the problem of detecting objects in frequently changing background, a method of constructing background model is proposed. Meanwhile, an improved Inter-frame Difference method is given to tackle the problem that moving objects cannot be wholly detected with Inter-frame Difference method. To overcome the drawbacks of optical flow method, i.e., the calculation involved is huge while there is time delay when detecting, an improved optical flow method which integrates with inter-frame difference method, is given.2. To deal with the problem of detecting moving objects in nonstationary complex environments, research is conducted regarding the relationship between inter-frame color co-occurrences and frequent foreground and background change. Based on the observation that inter-frame color co-occurrences are much significant for frequent changes in background than in foreground, a Bayes decision rule for classification of background and foreground changes based on inter-frame color co-occurrence statistics is derived. Methods to update background model in nonstationary environment are studied, and a way to maintain background model by updating reference background image and updating color co-occurrences is presented. Short-term and long-term strategies of updating the frequently changing background model are proposed. Based on the studies above, a novel method is given, which is based on color co-occurrence and Bayes decision rule, to detect foreground objects.3. An improved method of region based multiple target tracking is proposed. First, matching function is established according to the motion continuity principle of frame sequence in motion images so that matching test based on the characteristics of motion target zone can be conducted in the matching process. Second, Kalman filtering method is used to predict the next position of the target in the target zone so that the matching scope is narrowed and the searching speed is accelerated. Next, tracking control list is set up to record the newest motion parameter in the target zone to ensure the continuity of motion target tracking. Then, the missing problem oftarget caused by repeated record and pause is solved by exploiting centroid distance and delay searching. Finally, entering zone and leaving zone are set to judge the appearance of new target and disappearance of old target. Through all the above, correct tracking of targets is realized.4. A method based on Support Vector Machine to classify traffic objects is proposed. In order to classify the three kinds of traffic objects, that is, people, small vehicles, and large vehicles, there are three SVM classifiers to be built respectively. After training these classifiers with those samples of three kinds of traffic objects respectively, these classifiers can be used to classify the traffic objects. The results of experiments show that people, small vehicles and large vehicles can be classified correctly with the method proposed.5. Theories of video database management are studies. Traffic video information management system is designed, incorporating the relevant theories of video database management with the characteristics of traffic video information and the need for management of the information. The system can structuralize unstructured traffic video data, that is, to segment traffic video information into of shots, according to the video content, and then extract one or several key frames from each segment to represent this video segment. When the video data is stored, key frames and video segment will be stored separately. Index of key frames can be stored according to their content, so that the storage and search of the traffic video information can be conducted according to the content of the information mainly by identifying the characteristics of the target, e.g. color, texture and shape, etc.6. Based on the researches above, the system of "Regional Traffic Control Visualized Data Processing Platform" is designed and realized . Three application systems, which are based on this platform, are introduced. Research in the relevant technology is also conducted. In the plate recognition system, a method to locate plate by video detecting and color information is proposed. With this method, the plate regional image is captured directly with video detection technology and then the plate image is located and retrieved accurately with plate color information. In traffic violation detecting system, a method to distinguish violation is presented; in traffic video detecting system, methods to improve detecting accuracy and measures to accelerate the speed of detecting are also worked out.The main research work of this paper is supported by the Intelligence transportation Project of the Ministry of Public Security, "Research on Regional Traffic Control Visualized Data Processing Platform" (Intelligent Transportation theme: 20036152201) and the Project of Tackling Key Problems in Science and Technology of Xi’an, "Regional Transportation Video Intelligence Control Management System" (Intelligent Transportation video technology sub-item: GG04020).更多还原