【作者】 朱胜利；

【导师】 朱善安；

【作者基本信息】 浙江大学 ， 控制理论与控制工程， 2006， 博士

【摘要】 在视觉跟踪领域,Mean Shift算法是一个非常优秀的算法,是国外最近几年才发展起来的。国外研究者比较多,国内的很少,去年才见有几篇文章发表。初入跟踪领域,走过了很多弯路后,后来在Mean Shift算法上找到了突破口,并在该算法的研究中投入大量的时间,是以论文的题目以Mean Shift命名。Mean Shift在跟踪领域有很多很好的性质,比如实时性好、对遮挡、目标变形鲁棒性好等,但它也有一些缺点。作者针对性地对这些缺陷做了较好的改进。论文虽以Mean Shift命名,但涉及内容已超出Mean Shift的范畴。 第三章针对Mean Shift算法不能跟踪快速目标的特点,提出Mean Shift和卡尔曼滤波器相结合的算法,卡尔曼滤波器预测目标在本帧的可能位置,Mean Shift算法在该位置邻域内搜索,算法对快速运动的目标的跟踪效果良好,而且对遮挡问题也有很好的处理。 第四章提出Mean Shift的模板更新算法,Mean Shift没有模板更新的能力。模板更新对目标的跟踪至关重要,但没有通用的模板更新算法,一般都是针对某种跟踪算法如何解决模板更新的问题。本文提出一种基于卡尔曼滤波器组的Mean Shift模板更新算法。模板的元素取自目标特征值的概率,通过48个卡尔曼滤波器可以跟踪所有特征值的概率变化。算法构造巧妙,由于使用了较少的卡尔曼滤波器,算法实时性好,鲁棒性更好。 第五章提出核直方图的粒子滤波器目标跟踪算法。算法的系统动态模型具有对速度的学习能力,这样可以减少粒子的维数和所需要的粒子数。观测模型以Mean Shift算法对目标特征值的描述为基础,本文设计了一种新的模板更新算法,更新算法充分利用了粒子滤波器计算的中间值,因而没有增加算法复杂性。模板更新能够使得观测值的获得更可靠,因而提高了算法的鲁棒性。 第六章提出基于Mean Shift粒子滤波器的算法。粒子滤波器算法的主要缺点是需要大量的粒子来近似描述目标的状态,使得算法非常费时。新算法利用Mean Shift算法在重采样之前将粒子收敛到集合靠近目标真实状态的区域内,由于每更多还原

【Abstract】 Mean Shift is a very good algorithm in visual target tracking area. Many scholars in foreign countries has developed this algorithm in recent years, however, few scholars study it in our country. A few articles can be found last year ever.I experienced many difficulties, when I entered the realm of target tracking. Finally I found the Mean Shift algorithm. I have spent many times in this algorithm, so my dissertation is named Mean Shift. The algorithm has many advantages, for example: very good real-time, robust for occlusion and target distortion, however, it has a few defects. Many improvements are done for these defects. My dissertation is named Mean Shift although, but the content is out the scope of Mean Shift.In third chapter, an algorithm combining kalman filter and Mean Shift is bring forward, arm at the Mean Shift couldn’t tracking the fast moving target. Kalman filter is used to forecast possible position of target, then Mean Shift search the real position near the possible position. The algorithm has good effect to fast moving target, and can deal well to occlution.In fourth chapter, a template update algorithm of Mean Shift is put forward. Template update is very important to target tracking. There isn’t general algorithm for it however, many template update is arm to given tracking algorithm. In this chapter, an algorithm of template update of Mean Shift based a group kalman filter is proposed. The element of template is probability of eigenvalue of target. These probability are acquire by a kalman filter group which had 48 kalman filters. The whole algorithm conformation is very skilled. Because the number of kalman filter is lesser, algorithm is real-time, and robust.In fifth chapter, an algorithm based on kernel histogram particle filter proposed. System dynamic model of algorithm can learn the velocity of target, so the dimension of particle is reduced and the required particle is very few. Observation model of algorithm is based on Mean Shift describing the eigenvalue. In this chapter, a new template update algorithm is devised. New algorithm makes the best of middle value of particle filter, so that thecomplexity of algorithm is not added. Template update enable to get more credible observation, so improve the robustness of algorithm.In sixth chapter, an algorithm based on Mean Shift particle filter is proposed. The main disadvantage of particle filter is it require so many particles to approximately describing state of target, that the real-time of algorithm is not ensured. New algorithm use Mean Shift algorithm to muster the particle to area of real state after re-sample .Because the particle describe the state of more rationally , the needed particle is reduced, and the real-time of algorithm is improved. In this chapter, in the basis of kernel particle filter, the new algorithm could adjust number of particle adaptively, the agility of whole algorithm is better. Experiment result show algorithm could tracking target very well when target is occluded, and algorithm is more real-time comparing with traditional particle filter更多还原