【作者】 陈鹏；

【导师】 朱淼良；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2010， 博士

【摘要】 计算机视觉在最近几年中得到了很大的发展。然而计算机视觉领域仍有许多问题有待解决,其中背景减除与目标跟踪是目前研究者关注的焦点。背景减除和目标跟踪算法是视频监控、人机交互以及车辆导航等许多应用的关键技术之一。近年来,人们已经在背景减除和目标跟踪方面取得重要进展,提出了许多重要的背景减除以及目标跟踪算法。然而,背景减除和目标跟踪是一个非常困难的问题,更加鲁棒可靠的背景减除以及目标跟踪算法仍有待进一步发掘。本文对常见的背景减除及目标跟踪算法进行了研究与改进,提出了几种新的背景减除与目标跟踪算法。在背景减除中,混合高斯模型是一种较常见的模型。由于标准的EM算法是离线算法,因此虽然EM算法可以较为准确的估计GMM模型的参数,却无法用于在线背景减除应用中。为解决这个问题,本文提出了一种新的基于充分统计量的在线EM算法。实验结果表明这种算法可以高效准确的估计出GMM模型的参数。K-Means算法属于一种聚类算法,具有实现简单,运行效率高的特点。广义K-Means算法是对K-Means算法的扩展,能够估计出GMM模型中的所有参数。通过分析EM算法和K-Means算法之间的联系,本文提出了一种基于充分统计量的在线K-Means算法。实验显示这种算法的性能同基于Robbin-Monro近似的在线K-Means算法性能接近。本文基于以上算法对视频进行了背景减除实验,结果表明所提出的在线EM算法和在线K-Means算法都能够可靠的检测到运动目标。为了对算法的性能进行详细的分析和比较,本文采用一维及二维仿真数据对上述算法的性能进行了实验对比。高斯粒子滤波是贝叶斯滤波的一种半参数化的实现,已经被人们成功的用于目标跟踪中。本文提出了一种新的采样算法。通过将高斯粒子滤波的采样及预测步骤合并为一步,高斯粒子滤波的实现得到了简化。本文证明当系统的预测模型为线性时,新的采样步骤同原采样及预测步骤等价。当系统的预测模型为非线性时,可以先通过一阶泰勒展开对非线性模型线性化,然后再应用本文的采样方法。基于模板匹配算法的目标跟踪中,在遮挡或目标变形情况下,会出现跟踪不稳定的现象。为解决这个问题,本文提出了一种新的似然值度量函数。对比实验表明,改进后的模板匹配算法可以稳定的跟踪到运动目标最后本文在背景减除的基础上做了目标跟踪实验。实验结果表明背景减除提高了目标跟踪的稳定性。更多还原

【Abstract】 Although significant progress has been made in the field of computer vision, a lot of important problems in this field still need to be resolved. Among them, background subtraction and object tacking has received many researcher’s attention in recent years. Background subtraction and object tracking is one of the fundamental tasks in many ap-plications, such as visual surveillance, human computer interaction and vehicle navigation. Significant progress has been made in background subtraction and object tracking during the last few years, and several important algorithms have been developed for solving this problem. However, due to the fact that background subtraction and object tracking is a challenging problem, more efficient and robust background subtraction and object track-ing algorithm is still need to be explored. In this dissertation, some popular algorithms in background subtraction and object tracking are studied and several new algorithms are proposed.Gaussian Mixture Models is one of the most popular models for background sub-traction. Although EM algorithm can estimate GMM parameters accurately, it cannot be applied to background subtraction directly. The main reason is that EM is not an incre-mental algorithm. To solve this problem, a new sufficient statistic vector based online EM algorithm is proposed. Experimental results demonstrate that our approach can effectively and efficiently estimate GMM parameters.K-Means algorithm is a cluster algorithm which is simple to implement and efficient to run. Elliptical K-Means is an extension of K-Means and can also be used to calculate GMM parameters. By exploring the close relationship between EM and K-Means, a sufficient statistic vector based online K-Means algorithm is presented. Experimental results show that the performance of the new online K-Means algorithm is comparable with that of the Robbins-Monro approximated online K-Means algorithm. The usefulness of the proposed two online algorithms in background subtraction is validated on real video data and the performance of them is analyzed using simulated data.Gaussian Particle filter is a semi-parameter realization of Bayes filter and has been successfully applied to object tracking application. To simplify the realization of GPF, we proposed a new particle sampling method which combines the sampling step with state prediction step by taking advantage of the Gaussian assumption and by exploring the linear structure of the system dynamic model. We proved that when the system dynamic model is linear, the new sampling method is equal to the combination of the original sampling step and the state prediction step. When the system dynamic model is nonlinear, we use Taylor method to approximate the nonlinear model with a linear model. The proposed sampling method can then be applied to the approximated linear model.To solve the stability problem of template based object tracking algorithm, we pro-posed a new similarity measure function and applied it to an adaptive template matching based object tracking algorithm. Comparative experiments demonstrate that the improved algorithm can track object stably even serious occlusion present.Finally we demonstrate that the performance of object tracking can be improved greatly by combining background subtraction as a preprocessing step.更多还原