【作者】 廖雪超；

【导师】 甘朝晖；

【作者基本信息】 武汉科技大学 ， 模式识别与智能系统， 2006， 硕士

【摘要】 多目标跟踪技术是计算机视觉研究领域中最活跃的研究课题之一。随着现代计算机和信息技术的飞速发展及图形识别算法的革命性改进,多目标的实时追踪技术脱颖而出,在军事国防、交通监视、天文预测、智能监控等领域有着非常重要的实用价值。粒子滤波作为一种基于贝叶斯估计的非线性滤波算法,在处理非线性运动目标跟踪问题方面有独到的优势。但粒子滤波的框架中并没有包含数据关联的机制,当对多个目标进行跟踪,目标数发生变化或目标相互遮挡时,会出现跟踪目标的丢失。此外,多个目标之间的干扰也会影响跟踪的准确性。本文在应用粒子滤波进行目标跟踪的基础上,提出了将粒子滤波和背景建模相结合来处理多目标跟踪问题的技术框架,将多目标跟踪分为背景建模、多目标跟踪、初始化、重新初始化和粒子滤波五个模块。本文的研究成果如下:使用高斯混合模型(GMM)对图像的每个像素进行建模,来计算当前图像中属于背景像素的概率,从而提取出前景运动目标。在处理前景目标的阴影问题时,为了克服了图像局部或整体的光照变化的影响,通过建立亮度变化和色度变化的彩色模型,来区分阴影背景、高亮背景、普通背景及前景目标,然后自动选择阈值对图像的像素点进行分类。实验表明,该算法能够较好的区分前景目标及其产生的阴影,从而提高后续跟踪算法的效率。在背景建模的基础上,详细研究了对每个目标生成其粒子集,进行粒子滤波的算法流程及技术框架,实现了3种粒子重采样算法(替换选择算法、线性时间算法和权值函数算法)。在评价粒子权值时,为了较好的区分目标之间的颜色特征,对原有的用Bhattacharyya距离评价的算法进行了改进,仅统计背景建模后粒子区域前景像素的颜色分布,从而提高了目标跟踪的精度和效率。利用背景建模的结果,并结合目标运动的时间和空间关系,使用信任理论来处理新目标出现和旧目标消失的情况。实验表明,本算法能实现多个运动行人的有效跟踪。更多还原

【Abstract】 Multi-target tracking (MTT) is one of the most active research subjects in the area of computer vision research. With the rapid development of modern computer science and information technology, and with the marvelous renovation of image identification, MTT comes out into the open and is of great practical value in the field of military defence, medical research, traffic monitoring, astronomical prediction, intelligent supervision etc.Particle Filter, as a nonlinear filtering based on Bayesian estimate, has an advantage in the field of nonlinear moving target tracking. However, the framework of particle filter does not cover the mechanism of data association. When the target number changes or the targets shelter from each other during the process of multi-target tracking, target tracking will fail. Furthermore, the interference among some targets will influence the accuracy of tracking.Based on implementing target tracking by means of particle filtering, a technique framework of tracking target by integrating particle filtering and background modeling is presented. The multi-target tracking is classified into 5 modules as background modeling, multi-target tracking, initializing, re-initializing and particle filtering. The research results are as follows:The author models each pixel of the image with Gaussian Mixture Model (GMM), to calculate the probability of background pixel in the current image so as to abstract foreground moving objects. When processing the shadow of foreground moving objects, in order to overcome the influence of illumination change of part or integrity in the image, the color model of brightness distortion and chrominance distortion is established to distinguish shaded background, highlight background, original background and foreground objects, and then select the threshold automatically to classify the pixel points in the image. The experiments indicate that this algorithm can distinguish foreground objects and its shadow and increase the efficiency of tracking algorithm.Based on the background modeling, this article detailed investigates the algorithm flow and technique framework of generating the particle set of each object and particle filtering. Three particle resembling algorithm (select with replacement, linear time resembling, select with weight function) are realized. In the process of evaluating particle weight, in order to distinguish the different color features of the objects, the original algorithm (evaluating through Bhattacharyya distance) was improved. Only the color distribution of the foreground pixel in particle area after the background modeling is counted, therefore the accuracy and efficiency of target tracking are increased. By means of the results of background modeling, the connection between time and space of the moving targets, and belief theory, the problem of new target appearing and old targets disappearing can be solved. The experiments prove that this algorithm can realize the effective tracking of multiple moving persons.更多还原