【作者】 蒋良卫；

【导师】 桑农；

【作者基本信息】 华中科技大学 ， 控制科学与工程， 2013， 博士

【摘要】 目标跟踪一直以来都是计算机视觉领域中研究的难点跟热点。由于光照变化、背景杂乱、视角变化、遮挡以及目标形变等因素的影响,当前大部分的目标跟踪算法仍然很难达到理想的跟踪效果,距离实际的应用还有一定的距离。然而目标跟踪又是高层计算机视觉任务,如越界、翻越等异常行为检测、道路车流量和公共场所人流量统计、事件分析和事件理解的基础,因此,对于该课题的研究仍然具有十分重要的理论意义和现实意义。本文主要围绕均值漂移这种简单高效的产生式目标跟踪算法和现阶段研究比较热门的基于判别式模型的目标跟踪算法进行研究。主要研究成果如下：首先,基于背景加权和运动方向信息,我们提出了一种产生式目标跟踪算法。针对传统均值漂移算法目标模板可能包含一部分背景信息,本文用背景加权直方图来减少背景信息对跟踪模板的影响,并重点分析了它的作用原理。基于分析所得结论,我们用局部运动滤波器来估计目标的运动方向,然后通过相同的方式将跟踪目标的运动方向信息嵌入到均值漂移目标跟踪算法中,以提高跟踪算法的性能。此外,针对传统的均值漂移跟踪算法不能适应目标外观的变化这个缺点,我们提出了一个新的目标模板在线更新策略。对于背景模型的更新,不但用到了背景本身的信息而且还用到了待跟踪目标的信息。在真实的图像序列上的对比实验结果表明,本文提出的算法提高了原始算法的跟踪性能。其次,基于空间信息和模型更新,我们提出了一种判别式目标跟踪算法。在基于图像块袋的判别式目标外观模型中,针对图像块和目标本身的空间位置依赖关系以及图像块与图像块之间的上下文关系,在建立判别式模型和构造置信度图的时候,我们将空间和上下文信息以加权的方式融合到判别式跟踪算法中,使得所建立的目标外观模型更加科学合理、描述能力更强,置信度图更加精确。此外,我们提出了一种新的在线模型更新策略来适应不断变化的目标外观。与现有的几个比较流行的跟踪算法进行了比较,实验结果证明我们提出的方法提高了目标跟踪的精度和稳定性,相对于产生式目标跟踪算法具有较好的区分目标跟背景的能力。然后,基于判别式目标模型区分目标跟踪背景的能力,我们首先提出了一种判别式均值漂移目标跟踪算法。我们利用判别式模型区分目标跟周围背景的能力,将其嵌入到产生式目标跟踪算法-均值漂移目标跟踪算法中,以此来提高均值漂移目标跟踪算法区分目标跟周围背景的能力,从而提高跟踪性能,改善跟踪效果。此外,针对目标运动过程中的尺度变化,我们基于随机厥理论并结合产生式和判别式目标模型,提出了一种尺度自适应目标跟踪算法。跟几个主流的尺度处理目标跟踪算法的对比实验显示,该算法在跟踪变尺度目标方面具有一定的有效性,它提高了原始算法跟踪变尺度目标的能力。最后,我们对我们在本文中所做的工作进行了归纳总结,给出了本文的主要创新点,并基于本文的研究内容,探讨了未来在目标跟踪方面仍然可以研究的内容。更多还原

【Abstract】 Target tracking has been one of the focuses and difficulties in computer vision. Due to the illumination changing, background clutter, view changing, occlusion and deformation, at the present stage, most of these the tracking algorithms can not meet the requirement of application. However, the object tracking is the base of the high level computer vision task, such as abnormal behavior like crossing detection, people counting in public places and cars counting on the roads, event analysis and event understand. Therefore, research on this issue has important theoretical significance and practical significance.In this paper, we start our research beginning with the mean shift object tracking algorithm as well as these tracking approaches based on the discriminative object model The mean shift object tracking is a simple and effective algorithm which has been studied by lots of researchers. The tracking methods based on discriminative model have become a hot issue due to the development of the machine learning technology. The major results of our research are as follow:Firstly, we proposed an improved generative tracking algorithm based on the back-ground weighted histogram and the motion direction information. Since the object rect-angle used in the mean shift object tracking maybe include some background pixels, this paper utilizes the background weighted histogram to reduce the influence of the background information. We analyze its principle at details. Based on the analysis, we estimate the motion direction of the object using the motion filters. After that the motion direction in-formation is embedded into the mean shift tracking algorithm to improve the performance of the tracking approach. In addition, we propose a new method for model updating in or-der to track the objects whose appearance change continually. As to the background model updating, we utilize not only the background information, but also we make full use of the target object information. The experiment results show that our proposed method improve the performance of the original method.Secondly, based on the spatial information and model updating we proposed a discrim- inative object tracking algorithm. In the tracking algorithm based on bag of patches, in consideration of the spatial position dependence between these small image patches and the target object as well as the spatial context between these small images patches. We make full use of them when we build the appearance model and construct the confidence map so that the appearance is more scientific and reasonable and the confidence map is more accurate. In addition, a new appearance model updating strategy is proposed to adapt to the variance of the object appearance. The experiment results show that the improved tracking method enhances the performance in accuracy and stability. Comparing with the generative method, the discriminative method does better in distinguishing the object from the background.After that, we proposed a discriminative mean shift tracking method based on the good performance of the discriminative method in distinguishing the target object from the back-ground. We embed the discriminative model into the mean shift tracking framework and make use of its discriminative ability to separate the object from the background. In ad-dition, as to the scale change during the process, we propose a scale adaptive tracking al-gorithm based on random ferns classifiers as well as generative and discriminative model. The compared experiment results illustrate that the proposed method improves the ability of tracking object with scale change.Finally, we summarize our work in this paper, present its innovative points, and then we discuss the future work and research.更多还原