【作者】 王盼；

【导师】 赵泽茂； 郭春生；

【作者基本信息】 杭州电子科技大学 ， 信号与信息处理， 2009， 硕士

【摘要】 基于静止摄像机的运动目标检测与跟踪技术是计算机视觉及相关领域的基础问题之一,良好的检测跟踪算法能够为相关领域的高层应用如智能视频监控、模式识别、智能视频会议等提供高效有力的支持。但由于现实环境的复杂性和多变性,以及图像传感器分辨率的限制,导致复杂背景环境下运动目标的精确检测与跟踪存在一定的困难。基于上述原因,本文着重研究了视频中运动目标检测与跟踪技术,在总结前人的基础上提出了几种新的算法。论文第一章介绍了当前国内外的研究现状,第二、三章分别提出了两种新的运动目标检测算法,第一种是基于自组织映射的区域高斯模型的运动目标检测算法,第二种是基于自适应图切的运动目标检测算法。在论文的第四章还详细介绍了一种基于Rao-Blackwellized粒子滤波数据关联的多目标跟踪算法。另外在第五章总结了全文并对下一步的研究工作做出了展望。第二至四章所述算法的介绍如下:基于自组织映射的区域高斯模型的运动目标检测算法,采用单高斯模型和区域高斯模型的级联式检测方法,并在区域高斯模型的更新过程中引入了自组织映射的“竞争、合作”机制。算法不仅能够在水波纹和树叶晃动等复杂背景环境中检测运动目标,并且能够保证检测运动目标的相似性。使用PETS2002和Water Surface复杂背景图像序列的仿真结果验证了算法的有效性。基于自适应图切的运动目标检测算法,通过引入运动目标像素点数和前景背景邻接像素对数的卡尔曼预测和节点流量的自适应更新,成功将图切算法应用到视频图像的运动目标分割中,实现视频运动目标的连续全局优化分割。实验结果表明,算法在复杂背景条件下定量检测性能指标表现良好。基于Rao-Blackwellized粒子滤波器数据关联的多目标跟踪算法,该算法与卡尔曼预测相结合,比普通的粒子滤波器算法在跟踪精度上有较大的提高,并且减少了所需要的粒子数量。同时算法还可处理未知个数目标的消失、相互遮蔽、出现等的情况。最后给出了模拟场景的仿真跟踪结果。本文提出的两种运动目标检测算法与单高斯、多高斯及其它一些算法相比,在检测精度(如检出率、正检率等指标)上有了较大幅度的提升,对背景噪声的处理能力也更加优越。第四章所述多目标跟踪算法能够在使用较少数量粒子的情况下稳健的跟踪未知数目的运动目标,提高了跟踪速度与精度。更多还原

【Abstract】 Moving objects detection and tracking base on static video is one of the fundamental issues in the computer vision and other relative fields, algorithm which has excellent performance can give powerful support to high level applications such as intelligent video surveillance, pattern recognition, and intelligent video conference and so on. But due to the complexity and mutability of real environment, and also because of the limitation of image sensor’s resolution, resulting in that lots of difficulties exist in detecting and tracking moving objects in complex background situation.For the above reasons, moving objects detection and tracking technologies in video are researched in this paper, and several new algorithms are put forward on the basis of summarizing forefathers. The state of the art at home and abroad is presented in the first chapter. Two novel moving objects detection algorithms are proposed in the second and third chapter of this paper respectively, the first is moving objects detection by region Gaussian model base on self-organization mapping, the second is research on adaptive graph-cut algorithm to video moving objects segmentation. The Rao-Blackwellized Monte Carlo data association for multiple targets tracking algorithm is introduced in the forth chapter. A summary and outlook of this paper is given in the fifth chapter. The algorithms described in the second to forth chapter are introduced as below.The moving objects detection by region Gaussian model base on self-organization mapping algorithm, combined single Gaussian model with region Gaussian model to make up a cascading detection method, and the“competition & cooperation”mechanism was introduced in the step of region Gaussian model updating. This method not only could detect moving objects efficiently at the scene of shaking leaves, waving water surface and other complex background, but also could guarantee the integrity of detected objects. Through the experiments of PETS2002 and Water Surface image sequences, the characteristic of proposed algorithm was validated in various complex background environments.The adaptive graph-cut algorithm to video moving objects segmentation, through the Kalman prediction of the number of moving objects pixels and foreground-background adjoining pixels, and adaptive update of the nodes flow, the graph-cut algorithm is successfully applied to video moving objects segmentation. It achieves to continuous global optimization segmentation of video moving object. Experimental results show that the quantitative detection indicators of this algorithm perform very well in complex background conditions.The Rao-Blackwellized Monte Carlo Data Association for Multiple Target Tracking algorithm couples with Kalman prediction, which reduces the particles and increases tracking precision outstandingly compares with general particle filter. This algorithm also can cope with disappearance, shelter, and appearance of unknown number targets. The tracking results of simulating scenes are given at the end.Comparing with other algorithms such as single Gaussian and multiple Gaussians, these two moving objects detection algorithms proposed in this paper get good performance in detection precision and ability in dealing with background noises. The multiple objects tracking algorithm described in the forth chapter can track unknown numbers moving objects with less amount particles, which improves tracking speed and precision outstandingly.更多还原