【作者】 初红霞；

【导师】 王科俊；

【作者基本信息】 哈尔滨工程大学 ， 模式识别与智能系统， 2012， 博士

【摘要】 视频图像序列目标跟踪是计算机视觉领域的重点研究课题之一,正日益广泛地应用到安全防范、智能视频监控、人体运动分析、智能交通管理、人机接口交互、军事机器人视觉等领域。目标跟踪是对视频图像序列中的运动目标利用模式识别、图像处理等相关技术进行处理和分析,找到所感兴趣的目标所处的位置,最终获得运动目标的运动参数,如目标质心、速度、运动轨迹等,为视频运动分析和场景分析提供一定的数据来源,进而实现对运动目标的行为理解,以完成更高一级的任务。视频中存在复杂的背景环境、目标姿态变化、目标相互遮挡和交错、光照和气候变化等众多干扰因素,这些均使目标跟踪成为计算机视觉研究领域的难题。尽管国内外的学者对该领域进行了广泛深入的研究,同时也提出了很多解决方法,但仍然有许多关键问题没有得到有效的处理,迫切需要更加成熟稳健的跟踪技术和方法。本文针对视频图像序列目标跟踪中受研究人员高度重视的粒子滤波和均值移动算法中的关键技术进行了研究。以解决均值移动和粒子滤波两种算法的缺点为主线,以对复杂场景条件下遮挡等问题的处理和多特征融合为辅线,重点研究了均值移动算法中目标的建模方法,粒子滤波目标跟踪中的退化问题、粒子滤波用于多目标跟踪等问题。论文的主要工作有：(1)基于空间相关背景加权直方图均值移动目标跟踪提出用空间相关背景加权直方图对目标进行建模,设计出了基于空间相关背景加权直方图的均值移动跟踪算法,首先给出新算法详细的推导和证明过程,以及用空间相关背景加权直方图如何描述目标外观。其次,提出背景模型的动态更新方法以提高背景环境有很大变化时跟踪的精确性,使跟踪不过分依赖于目标的初始定位。然后将特征融合思想引入均值移动算法中,进一步解决目标和背景或遮挡物颜色相近时出现的跟踪失败或误跟踪问题,同时克服了均值移动单一颜色特征缺乏空间信息和易受光照等外界环境影响等问题。最后在实际的视频序列数据中进行测试、分析和对比,以验证跟踪算法的鲁棒性。实验表明,与传统的均值移动跟踪方法和空间直方图均值移动跟踪相比,本文提出的算法在初始目标定位不准确的情况下,对具有较长时间以及和被跟踪目标颜色相近背景物的严重遮挡等问题上,均取得了较好的实验结果。(2)均值移动优化的粒子滤波目标跟踪针对粒子滤波器的退化问题,通过引入均值移动算法对粒子进行有效传播,各种优化机制整合在两种算法中实现对粒子有效分散和聚类,减少了粒子集,提高了粒子滤波的计算和采样效率。在粒子传播过程中,用具有大小和尺度自适应的均值移动优化每个粒子的位置和方向,初步解决退化问题。不确定性权值自适应调整算法使粒子自适应更新权值,权值更新中又结合多特征融合的有效方法,使目标外观模型得到了更好的描述,进一步使退化问题得到有效解决。为适应复杂背景环境,算法又辅以相应的模板更新策略。在具有相似背景、相似背景物的遮挡以及目标有较大尺度变化等不同跟踪条件的视频数据上进行测试,实验表明,这种多特征融合的均值移动优化粒子滤波算法与现有的粒子滤波方法相比,退化问题得到了有效解决,取得了明显的改进。(3) Adaboost检测和混合粒子滤波相融合的多目标跟踪针对粒子滤波多目标跟踪时如何连续维持目标分配的多种模态,以及如何控制多模式的增长问题,设计了非参数化递归模型的混合粒子滤波。算法能较好的保持和有效处理多模式问题,在标准粒子滤波失效的地方保持固有的多模态,有利的解决了在非约束性跟踪应用中的很多难题。首先给出由蒙特卡罗推导递归实现混合粒子滤波的过程,通过混合权值的计算实现粒子间的相互关联。其次,在跟踪算法中,构造了由多特征相融合的动态模型和Adaboost检测信息合并成的混合观测似然函数,融入Adaboost建议密度后算法能快速的检测进入场景的目标而粒子滤波过程则能保持个体目标的有效跟踪,多特征的融合同时能有效的估计目标外观有较大变化的样本。第三,为了克服模型漂移现象,提出使用交换概率主成分分析的模板更新方法,使目标在背景复杂多变时稳定跟踪。在刚性、非刚性以及数目变化的多目标视频序列中对算法进行测试,实验表明,算法对数目确定或数目变化的多目标能实现有效跟踪。(4)改进的粒子滤波多目标跟踪针对粒子滤波多目标跟踪中数据关联和估计问题,在粒子滤波和Gibbs采样框架下对多目标进行研究。首先在给定几个观测过程时把经典粒子滤波扩展成多目标状态过程的估计。然后从随机性这个角度考虑数据关联,用Gibbs采样作为估计和分配关联向量的主要方法,即通过粒子集表示目标状态的联合后验分布,目标状态向量和关联概率被联合估计不用经过列举,修剪、门限等运算,避免了合并的弊端。在纯方位目标和实际的视频序列中对算法进行测试。实验表明,算法有较强的解决数据关联问题的能力,即使在密集杂波的干扰和非线性下,也有另人满意的性能。更多还原

【Abstract】 Target tracking Video Image Sequence is one of the key research subjects in computer vision field and now is extensively used in Security guard, intelligent video surveillance, human movement analysis, intelligent traffic management, Man-machine interface interaction and military robots vision and so on. Target tracking analyzes and processes the moving target in video image sequence by Pattern recognition and image processing technologies so as to find the interested target location, and then ultimately get the parameter of the motion target, including target centroid, velocity and trajectory. These provide a source of data for video motion and scene analysis, thus achieving behavior understanding towards moving object so as to complete a higher-level task. There are many interference factors in videos, including complicated background environment, target posture change, target frequent occlusions and cross, illumination and climate change, etc. All these make target tracking a difficulty in computer vision research field. Though domestic and foreign scholars had extensive and deep research in this field, as well as solutions, many key problems haven’t effective treatment. All these urgently need mature and steady tracking technologies and methods.This dissertation studies key technologies in particle filter and mean shift, which is paid more attention by researchers in video image sequence tracking. Regard solving the defects of mean shift and particle filter algorithms as the main line and treating occlusions in complicated backgrounds as well as multi-features integration as the auxiliary line. The paper focuses on resolving modeling method in mean shift, degradations and multi-targets tracking in particle filtering target tracking, etc. The main tasks in this thesis are:(1) Multi-features integration on mean shift target trackingMethod of targets modeling with space corrected background weighted histogram is provided. Mean shift tracking is designed based on space corrected background weighted histogram. Firstly, this thesis offers the detailed derivation and proof procedure of the new algorithm, as well as how to describe the targe appearance with space related background weighted histogram. Secondly, this thesis proposes the dynamic updating methods of background model to improve the tracking accuracy when background environment have great changes, so that tracking will less depends on target initial position. Then, feature integration is introduced into mean shift tracking to further solve tracking failure or false when there is similar color of target and background or blocking matters. Finally, testing, analysis and contrast is running in actual video sequence data so as to verify the robustness of tracking algorithm. Comparing with traditional mean shift and spatiogram target tracking, experiments shows that the mentioned algorithm achieves good results in the case of inaccurate initial target positioning and severe occlusions with long-term and colors similar to being tracked targets.(2) Multi-features integration particle filter target tracking optimized by Mean ShiftFocusing on degradation issue of particle filter, particles are effectively spreaded through introducing mean shift algorithm. All kinds of optimized mechanism is integrated in the two algorithms to achieve effective scattering and gathering for particles, so as to reduce particle sets and improve the calculating and sampling efficiency of particle filtering. In the period of particle dissemination, positions and directions of each particle is optimized by using sizes and scales adaptive mean shift, to initially solve the degradation issues. Uncertainty weight adaptive adjustable method makes particle adaptive updating weight. Effective methods of multi-feature integration are combined with weight updating so as to better descripe target appearance model, and effectively solve future degradation issues. In order to adapt to complicated background environment, algorithms combined with the corresponding model updating measures. Algorithm is test on different tracking video data which have similar background, similar background occlusions as well as large scale change. The experiment shows that comparing with proposed algorithm and existing particle filtering method, degradation issue is well solved as well as achieving significant improvement.(3) Mixture particle filtering multi-targets tracking based on multi-features integrationFocus on the problems how to continuously maintain multi-model of target distribution in particle filtering multi-target tracking, as well as how to control the multi-model increasing, this thesis designed mixture particle filter of non-parameterized recursive model. It can effectively maintain and treat multi-model issues. Inherent multi-model is maintained when standard particle filtering is ineffective and many difficulties are effectively solved in non-constraints tracking applications. Firstly, the mixture particle filtering process is recursive realized by Monte Carlo derivation. Correlation between particles is achieved by mixture weight algorithms. Secondly, in the process of tracking algorithm, it constructs mixture observation likelihood function by mixing the dynamic model of multi-features integration and adaboost detecting information. Suggestion density integrated adaboost can rapidly detect the targets, and particle filtering process can maintain single target effectively tracking. Meanwhile multi-features integration can effectively estimate the samples when target appearance has greatly changes. Thirdly, in order to conquer model drifting phenomenon, this thesis proposes methods of switching probability principal component analysis to update model, so that the target can steadily track when backgrounds are complicated and changeable. Test towards algorithms in rigidity and non-rigidity and change quantity multi-targets video sequences, and proofs that algorithms achieves effectively track towards determinate number or changeable number multi-targets.(4) Multi-target tracking on multi-features integration particle filteringFocus on data association and estimation problems in particle filtering multi-targets tracking, this thesis makes researches the multi-targets in the framework of particle filtering and Gibbs sampling. Firstly, it extended classical particle filtering into multi-targets state estimation in the given several observation process. Then data association is considered from randomness point of view. Gibbs sampling is regarded as the methods of estimation and distribution correlation vector, that is, Joint posterior distribution of target state is expressed through particle set. Target state vector and association probability was jointly estimated without list, trim, threshold and other algorithms. This avoids merger drawbacks. Test is running in bearings-only target and real video sequence. Experiments show that algorithms have strong ability of solving data association problems. There are satisfactory performance even in the circumstance of dense clutter interference and nonlinearity.更多还原