【作者】 刘涛；

【导师】 李明；

【作者基本信息】 兰州理工大学 ， 计算机应用技术， 2009， 硕士

【摘要】 目标跟踪是计算机视觉的一个重要分支,融合了图像处理、模式识别、人工智能、自动控制以及计算机应用技术等相关领域的先进技术和研究成果。实现目标跟踪的关键在于完整地分割目标,合理、快速地提取特征和准确地识别目标,同时要考虑算法实现的时间,保证实时性。在复杂环境下对目标进行跟踪是非常困难的,仅仅依靠一种或少数几种识别手段很难准确地进行目标识别,必须尽可能利用多个目标属性、多种方法,进行目标综合识别和跟踪。复杂场景中遮挡目标的跟踪一直是目标跟踪领域的困难问题。本文主要针对跟踪过程中的遮挡难题进行研究,详细论述了mean-shift、Kalman滤波、粒子滤波、交互式多模型等算法,在分析算法对遮挡处理的优缺点的基础上,对其进行综合有效地改进。论文的主要工作有:(1)提出了一种结合均值漂移和粒子滤波的跟踪算法:利用改进的均值漂移算法跟踪目标,根据目标搜索公式确定遮挡因子,判断遮挡程度;当目标被严重遮挡时,采用mean-shift筛选出权重粒子,建立更新模板,持续、稳健地跟踪目标。算法实时性强,克服了粒子滤波器的退化现象并有效缩短了计算时间,解决了跟踪过程中目标的部分遮挡和全遮挡问题。(2)提出了一种改进的基于Kalman滤波和粒子滤波的交互式多模型跟踪算法:利用卡尔曼滤波匹配系统线性部分,粒子滤波匹配非线性部分,根据匹配深度来判断目标受遮挡的程度,采用迭代的多级粒子滤波方法进行重采样,并结合卡尔曼滤波更新模型概率,解决跟踪过程中的严重遮挡问题。改进算法提高了模型滤波速度和目标状态的估计精度,缩短了计算时间。更多还原

【Abstract】 Object tracking is an important branch of computer vision,which combines advanced technologies and research achievements in image processing,pattern recognition,artificial intelligence,automatic control,computer application and other relative fields.It include segment target,pick-up character and recognize objects.In practical application,the operating time of tracking algorithm must be also considered.Object tracking in a complex environment is very difficult.Only depending on one or a very few identify means is difficult to recognize objects accurately,so it need to make full use of multiple objects property and several means to track.This thesis mainly aims to analyze and solve occlusion problem in object tracking.It detailedly discusses the occlusion handling methods in several algorithms such as mean shift,kalman filtering,particle filtering,interacting multiple model etc,and makes effective improvement.The following is done in this thesis:Firstly,an improved algorithm is proposed by combining the advantages of the mean shift particle filter in this paper.Generally,the improved mean-shift algorithm is used to track objects;then the factor is given to judge the degree of occlusions by the searching formula.When the serious occlusion exists,the weighted particles are selected by the mean-shift to establish the update template.The algorithm can track the objects continually and steadily after occlusion.Experimental results show that the real-time algorithm tracks objects satisfactorily.It solves partial occlusion and full occlusion.So the degeneracy problem is efficiently overcome and the computational cost is decreased.Secondly,an improved algorithm based on Kalman filter and particle filter is proposed,in which Kalman filter is used to match the linear part of the system and particle filter is used to match the non-linear part of the system,the degree of occlusion is determined according to the match extent when the serious occlusion exists,the iterative multistage particle filter is exploited for re-sampling,then combined with Kalman filtering to update the model probability which can track the objects continually and steadily.Experimental results show that the proposed algorithm meets the real-time requirement,improves the speed of the model filter and the estimated accuracy of the object state,and reduces the computing time effectively.It also solves the occlusion problem in the process of tracking.更多还原