【作者】 高韬；

【导师】 刘正光；

【作者基本信息】 天津大学 ， 检测技术与自动化装置， 2010， 博士

【摘要】 交通视频监控系统通过摄像机获取场景视频序列,以此对道路场景中的运动目标进行检测﹑定位和跟踪,并在此基础上分析和判断目标的行为,从而做到对违章行为的实时监控记录。由于其理论还不是很完善,新的方法和技术还有待开发,因此,对交通运动目标进行检测与跟踪,是一项既有理论意义又有实用价值的课题。本论文针对智能交通监控中的关键技术问题,所做的主要工作包括以下几个方面:1.结合了二进冗余离散小波变换的各子带系数高度相关、方向选择性、各子带与输入信号大小相等以及平移不变的性质,提出了在冗余离散小波变换域进行运动区域提取的算法。通过二进冗余离散小波变换直接在小波域提取运动区域从而检测运动目标,在一定程度上克服了传统时域检测法的缺陷。2.背景建模部分,采用Marr小波核函数的背景建模算法。通过对图像中的每个像素利用Marr小波核函数进行分布建模以及利用输入帧实时更新模型,从而可靠地处理了光照、混乱运动的干扰。3.对于阴影处理,针对路面阴影处具有稀少的边缘细节,而目标内部边缘细节丰富的特点,采用多尺度几何边缘识别算法,来区分阴影背景与前景目标,克服了传统阴影去除算法当目标和阴影在颜色信息上没有明显差别时常会误判以及对光照变化敏感的缺陷,从而准确有效的提取出前景运动目标。4.车辆跟踪部分,在运动检测的基础上,详细研究了利用SIFT (Scale Invariant Feature Transform)特征粒子滤波目标跟踪技术,通过保留特征性好的独立粒子缓解了粒子退化现象,并采用自适应均值滤波,获得目标边界准确位置。同时利用距离测度以及Bhattacharyya相关系数来处理新旧目标出现和标消失的问题。此外,采用队列链表法记录多运动目标之间的数据关联,在提高检测准确率的同时降低了运算的复杂度,从而提高了目标跟踪的精度和效率。5.交通视频监控硬件系统采用多摄像头,多分辨率,多视角,远近景相结合的工业控制计算机系统架构。通过远近景结合,既可发现违章目标,又能通过近景抓拍牌照信息,在提高交通执法监控效率的同时降低了设计成本。更多还原

【Abstract】 Traffic video monitoring system detects the position of moving object and tracks it according to the road scene obtained by camera; and then analyses the behavior to record violations for a real-time monitoring system. However, due to the short history of development, some important problems are still unresolved, and new methods or techniques are also needed. Thus, traffic video based moving objects detection and tracking is a subject with both theoretical and practical value. This dissertation dose the researches focused on the key technical problems about intelligent traffic monitoring, and the major works include as follows:1. As the coefficients of binary redundant discrete wavelet transforms in each sub band are highly relevant, direction selectivity, and the sub-band signal is with the same size of input signal, as well as translation invariance, a redundant discrete wavelet transforms domain based motion region extraction method is presented. The moving objects are directly detected in the wavelet domain which overcomes the defects of traditional time-domain detection methods.2. For background modeling, Marr wavelet kernel function is used in a probabilistic background modeling method. Each pixel of the background image is modeled by using Marr wavelet probabilistic distribution, and the real-time input frame is used to update the background model in order to reliably deal with the interference of chaotic movement in background.3. For the shadow elimination, it is generally believed that shadow on the road has scarce details while the internal region of moving object not. A multi-scale edge geometric recognition method is used to distinguish between background shadow and foreground, and then the threshold is automatically selected for image pixel classification. The method overcomes the traditional shadow elimination methods which are influenced with the light changing, and overcomes the misjudgment if object and shadow have the similar color information, so object segmentation is more accurate.4. For vehicles tracking, based on the identification of moving objects, a SIFT (Scale Invariant Feature Transform) features of particle filter tracking technology is used to overcome the degeneration, and a detail description of object tracking technology about feature extraction and object positioning is presented. Combined with the movement and space relations, an adaptive Mean Shift filter is used to obtain the exact location of object border, and Bhattacharyya trust theory is used for the correlation coefficient to deal with the emergence and disappearance of an object. In addition, a linked list queue data association is used to record the relation between moving objects for improving the detection accuracy and reducing the complexity of computing.5. The hardware system of traffic video monitoring contains multi-camera with multi-resolution and multi-perspective, which are combined with industrial control computer system. The system can not only find vehicles of violation, but also capture the information of vehicle license by a close-range camera, which improves the efficiency of traffic monitoring and reduces the design cost.更多还原