【作者】 张竞；

【导师】 王向军；

【作者基本信息】 天津大学 ， 测试计量技术及仪器， 2007， 硕士

【摘要】 目标检测与跟踪技术是计算机视觉的主要研究方向之一,它是智能监控、人机交互、移动机器人视觉导航、全景战车等应用的基础和关键技术。动态图像序列中,除了具有与图像一样的空间特性外(如颜色信息、纹理信息等),还具有时序特性,即动态图像序列中的运动信息。因此,与传统的图像分析技术相比,基于动态图像序列的运动信息分析在上述领域中起着更加重要的作用,然而在实际的应用中,由于摄像机抖动、运动目标的不规则运动以及摄像机运动所引起的背景运动等因素,使得运动目标的检测变得困难,本文对于动态背景下图像序列中的运动目标检测展开了研究,研究内容和主要成果如下:(1)针对动态背景下微小运动目标的检测,提出了基于帧间差分向量一阶范数检测技术与空域峰值点双窗口搜索技术相结合的检测方法,该方法不仅能检测出帧间位移不小于一个像元的点目标,而且能检测出多帧累积位移大于一个像元的点目标,实验结果表明,该方法能有效地检测出动态背景下的多个运动点目标。(2)提出了基于六参数仿射模型的全局运动估计补偿检测法,全局运动估计是指对图像序列中造成背景运动的摄像机运动进行估计。通过全局运动估计,可以获得帧间像素的相关性,用得到的全局运动估计参数对图像序列进行补偿,从而将复杂运动背景下的运动分析转化为静态背景下的运动。在本论文中,该算法对动态序列图像中的运动目标(如运动车辆、行人等)检测有不错的效果。(3)对上面全局运动估计中使用的宏块匹配算法进行了优化,从而降低了计算量,提高了程序的运行速度,增加了检测的可靠性。更多还原

【Abstract】 The technology of object detection and tracking is one of the hotspots in the field of computer vision, which is also the basic and important technology in the application of smart surveillance, human-machine interface, mobile Robert navigation, full-sight caterpillar system and so on. Dynamic image sequences possess not only spatial properties like images (such as color, texture etc), but also temporal properties, namely, the motion information. Thus, compared with the traditional image analysis techniques, motion information analysis in image sequences plays more important role in solving the above mentioned problems. However, in the practical applications, jittering camera, moving objects’irregular motion, the moving background caused by camera motion etc make the detection of moving objects difficult. A thorough research on the detection of moving objects in dynamic image sequences has been carried through in this dissertation. Main research works and achievements of this dissertation are listed as follows:(1) A method of detecting moving point objects in dynamic image sequences is proposed, which is based on the combination of two methods: detecting method based on first-order norm of frame difference vectors and the method based on peak value detection in airspace searched by double windows. It can detect not only the point object whose displacement is no less than one pixel between two continuous frames, but also the point object whose displacement is no more than one pixel between two continuous frames and whose displacement gathered of multiple continuous frames is larger than one pixel. The result indicates that, this method provides a new solution to detect multiple moving point objects in dynamic image sequences effectively.(2) A method is proposed to estimate global motion, which is based on 6-parameter affine model. The method is to estimate the law of the camera motion, which causes the background moving in the image sequences. By the estimation of global motion, the pixel correspondences between adjacent frames can be attained, and frames can be compensated with the parameters get by the global motion estimation, thus the problem with complex moving background can be simplified to the one with static background. This method has a fine effect in the detection of moving objects (such as moving cars, moving people etc) in dynamic image sequences. (3) The block matching algorithm used in the global motion estimation is optimized, which reduces the calculate amount, increase the speed of program running, and increase the reliability of moving objects detection to a great extent.更多还原