【作者】 刘英霞；

【导师】 王欣；

【作者基本信息】 山东大学 ， 信号与信息处理， 2007， 博士

【摘要】 对运动的人体进行实时跟踪就是通过对采集到的图像序列进行处理和分析,获得关于人体的相关信息。它在智能监控、高级人机界面以及基于运动的诊断与辨别等方面都有广泛应用和潜在的经济价值。因此该方向是近年来计算机视觉领域研究的热点问题之一,特别是在复杂的动态背景下如何完成准确地跟踪更是人们讨论的重点,并受到日益广泛的重视,而图像去噪是人体跟踪的重要组成部分之一。基于此,本文对在不同的背景环境下运动目标的跟踪问题进行了讨论和研究,主要包括图像小波域去噪、图像边缘增强、在小波域进行目标检测、复杂背景下基于贝叶斯模型的运动目标检测、目标分析、利用卡尔曼滤波理论进行目标跟踪等。人体的实时跟踪系统是对人体进行实时监控的智能系统,在本论文,对其中的具体问题进行了细致的分析和讨论,提出了一些新的算法。本文的主要贡献如下;1.在对包含有目标的图像进行分析前必须要进行图像的预处理,以尽量减少噪声对目标的干扰。因此,文中提出了不同的算法来完成该任务;(1)最佳软门限去噪;在Donoho软门限的理论基础上,对小波系数的收缩量进行进一步的讨论,得到一个最佳的软门限值完成图像的去噪。该方法利用最小均方误差准则,对原有的软门限值进行了优化,推导出在最小均方误差的条件下对于不同噪声分布的最佳去噪算法。理论和实例证明,该算法可以得到一个比原有软门限方法更好的滤波结果。(2)双Haar小波变换系数的MAP估计完成图像去噪;首先对小波基的选取进行了讨论,由于双Haar小波不仅具有良好的边缘检测性能而且有更好的平滑噪声的能力,因此适用于信号去噪。该方法给出了基于MAP的双Haar小波收缩算法,并对其在图像去噪中的应用进行了讨论。该算法充分利用了信号的双Haar小波两个变换系数间的相关特性,得到小波系数的最佳估计。这种小波收缩算法与现有的软门限方法相比较,可以给出更好的去噪结果。(3)自适应边缘增强算法;边缘增强是图像处理的一个重要研究领域,它主要用于对模糊图像的锐化。但是,现有的图像边缘增强方法对噪声敏感。由于中值滤波是一种用于图像去噪的非线性滤波器,为了提高中值滤波的细节保持能力,可采用多级中值滤波器。由于在图像的边缘部分,多级中值滤波和均值滤波的输出差较大,而在图像的平坦部分输出较小,因此该算法在传统的高通滤波器的基础上,利用多级中值滤波和均值滤波的差作为参数,提出了一种自适应图像边缘增强算法。它在增强图像边缘的同时,还具有抑制噪声的能力,因此具有良好的应用前景。2.在对运动人体进行跟踪时,准确地检测出目标是非常关键的一步,因此本文对不同的背景情况下的目标检测算法进行了详细的讨论,提出了以下几种改进算法;(1)基于贝叶斯准则水平投影优化的运动目标检测算法;在目标检测系统中,阈值的选取是一个非常关键的问题,它直接决定了检测结果的准确性。如果选取经验值作为门限将目标从背景中分离出来,有时并不能得到理想的检测结果。新算法充分考虑了仅仅依靠阈值来确定目标区域大小的局限性,结合贝叶斯判决准则,给出了一个动态阈值,利用该阈值对二值化图像的水平投影做进一步分析,在去噪声的同时保证目标的完整性。该方法当外界环境光线变化较大时,可以有效地改善光线变化带来的影响,从而使人体目标的检测更为准确。(2)基于滑动窗口的双Haar小波运动目标检测算法;该算法不再直接在时域对目标分析,而是从频域的角度出发,提出一种新的运动目标检测方法。在对运动目标检测之前,首先对差分图像去噪,然后对去噪后的图像进行边缘增强,使目标区域更为明显。最后,利用图像的能量特性得到差分图像的特征图像,区分背景和目标,并在特征图像直方图的基础上得到判断阈值,完成图像的二值化。实验证明,利用该方法可以得到更好的检测结果。(3)基于贝叶斯模型的动态背景检测;在跟踪目标时,有时会遇到复杂的动态背景,原有的检测方法有时不能得到满意的结果,因此提出一种新算法。该算法利用贝叶斯模型结合核密度函数理论,把阈值的选取放到概率空间上考虑,给出了在复杂的动态背景下检测出运动目标的方法。首先,在图像相关性的基础上,利用非参数核密度估计的方法,建立前景和背景的核密度函数,再利用贝叶斯理论,估计出背景和前景的先验概率,两者相结合得到一个估计的阈值,从而实现目标和背景的分类。3.在准确检测出目标后,就要对运动目标进行跟踪了。关于运动人体的跟踪本文提出了下面的改进算法;快速卡尔曼滤波算法在人体跟踪中的应用;在对运动目标进行跟踪时,为使跟踪准确无误地完成,需要根据目标当前的运动轨迹来预测下一时刻的位置并对该时刻的位置进行修正。卡尔曼滤波预测理论可以解决这类问题,将该理论应用到人体跟踪中,可以达到理想的跟踪效果。该算法在原有方法的基础上,首先根据运动目标的对称性,确定卡尔曼滤波的状态方程,然后对系统的增益方程和常数矩阵做了合理的假设,通过理论推导得到其常数值。实验表明,该算法不仅可以使目标的跟踪更为准确而且大大提高了计算的速度。通过对室内和室外采集的图像序列进行实验,表明综合利用上述方法能够准确检测出运动者的位置,完成运动目标的跟踪。更多还原

【Abstract】 The real-time tracking for moving people is to process and analyze the sampled image and to obtain the information about the body. The system has the wide use and economical worth in the smart surveillance and monitor, high-grade Human-Computer interface, motion-based diagnosing, etc. So, this field has been become one of the most active research area in computer vision recently. How to detect the moving object from the complex background becomes the significant discussing topic especially and extensive attention has paid to it. As we know, image de-noising is the important part of the image processing. So, the tracking problem is discussed under different background environment in this paper, it mainly contains image de-noising in wavelet domain, image edge enhancing, target detecting in wavelet domain, moving object detecting on complex background with Baysian rules, moving object analyzing and object tracking with Kalman filer, etc.The real-time tracking system for people is a general frame of real-time monitor, the concrete problems are analyzed and several new algorithms are proposed here. The main ideas are as follows:1. The image which contains target must be pre-processed in order to decrease the interference of the noises. So, the different algorithms are proposed in the paper:(1) Image denoising by the optimal soft threshold: the shrinkage of the wavelet coefficients are discussed further on the basis of Donoho theory and an optimal soft threshold is obtained to finish the image denoising. The MES rule is used to optimize the original soft threshold and the optimal denoising algorithm is deduced for different noises. The theory and experiment has proved that a better denoising result can be obtained than original method.(2) Image de-noising based on double Haar transform MAP estimate: At first, the wavelet basis selection is discussed. Double Haar transform not only has the better edge detecting property but also can smoothen the noises more effective. The new method gives a wavelet shrinkage algorithm based on MAP estimate and discusses its application in image de-noising. The correlation between the double Haar transform coefficients are used to obtain the optimal estimate. This shrinkage algorithm can get better de-noising result than the original soft threshold.(3) Self-adaptive edge enhancing algorithm: Edge enhancing is an important domain in image processing, and it is used to sharpen the blurred image. But, the present method is sensitive to noise. In order to raise the details protection abilities of the median filter we can use the multi-median filter. Because the output between the multi-median filter and mean filter at the image edge part is larger than the image flat part, the difference is used as the threshold to propose a self-adaptive edge enhancing algorithm. This method has the ability to constrain noises when the image edge is enhanced and it has the better application prospects.2. Detecting object correctly is a key step when the moving people is tracked, so, the new detecting algorithms are discussed under different circumstances in this paper and several improved algorithms are proposed:(1) Moving object detecting algorithm based on Baysian rules: In the object detecting system, the threshold selection is a key problem and it determines the accuracy of the detecting result. If the experience value is selected as the threshold to separate the object from the background, we can not obtain the ideal result sometimes. The new algorithm considers the limitation to define the object region by the experience value, and combines with the Baysian rulers, analyzes the horizontal of the binary image with a new dynamic threshold, ensures the object region complete and without noises. The method can improve the influence of changing light and make the detecting result more accurate.(2) Moving object detecting algorithm based on moving window double Haar transform: This mean does not analyze object in temporal domain any more, it starts from wavelet domain, and proposes a new algorithm. Before detecting object, the edge of de-noising difference image is enhanced and make the object region more obvious, then, the feature image is obtained with the energy character to differentiate background and object, and the judging threshold is obtain based on the histogram of the feature image to finish the image binarization. The experimental result shows that the new algorithm can get better detecting result.(3) The dynamic background detecting algorithm based on Baysian model: We usually meet complex background when we track the object. Baysian model is combined with kernel density function to give a new detecting object method in probability space. At first, the kernel density function of background and foreground is built based on the image correlation. Then, the prior probability of the background and foreground is estimated with Baysian theory. At last, a dynamic threshold is obtained to classify the target and background.3. We should track the moving object after we detect he (or she or them). The following algorithm is proposed in this paper for people tracking:Fast Kalman filter algorithm for moving object tracking: In order to track the moving object correctly, we should predict the position of the body in the next time according to the present frame and revised the position. Kalman filter can solve the problem and it can get the ideal tracking result. The new algorithm defines the state function at first based on the Kalman theory, then, deduces the value of the augment matrix and constant matrix. Experimental result shows that the method can not only track the object correctly but also improve computing speed.The experimental results of the indoor and outdoor circumstances show that the above-mentioned means can be used to detect and track moving people correctly.更多还原