【作者】 邓志辉；

【导师】 路林吉；

【作者基本信息】 上海交通大学 ， 控制理论与控制工程， 2010， 硕士

【摘要】 视觉监控中的多物体跟踪是计算机视觉研究领域的热点问题之一,尤其是近年来,视频监控系统发挥着越来越重要的作用,它广泛应用于民宅、停车场、公共场合、银行等一些场所的实时监控。本文基于为使用一个固定的普通彩色摄像头来监控户外或者室内场景场合,设计了一种综合运动检测与物体跟踪的智能视觉监控系统,对数目变化的多物体能自动完成检测和跟踪,并保存轨迹信息。多物体跟踪的难点在于跟踪目标变化不定、实际场景复杂、物体存在变形等,本文基于运动检测与物体跟踪相结合的思想,将多物体跟踪系统分为运动检测模块、团块检测模块、跟踪模块和轨迹产生模块四部分,对提高运动检测与物体跟踪的实时性、鲁棒性与精确性进行了研究。本文的主要研究内容如下:运动检测部分,首先详细分析了基于码本模型的背景差法,在原算法基础上,将像素在时域视为高斯分布,使之更符合统计规律和人体视觉系统,重新定义了码本、亮度失真度及更新规则,从而能检测更加完整的前景目标。接着详细分析了基于贝叶斯分类的背景差法,并基于码本模型中的颜色空间模型,设计了新的阈值化法,能够一定程度上抑制拖影现象。实验表明,这两种背景差法在存在环境噪声、运动的背景等情况下,都能有效地检测出运动的物体。目标跟踪部分,在运动检测的基础上,给出了多物体跟踪的框架,即将多物体跟踪分解为多个单物体跟踪的组合。一方面,将跟踪问题视为最优状态估计问题,详细研究了利用粒子滤波进行物体跟踪的算法流程和实现方法。并采用目标颜色特征和运动特征相结合的似然函数,以及采用MCMC改善粒子的重要性分布,从而提高了目标跟踪的精度和效率。另一方面,将跟踪问题视为一分类问题,详细研究了利用On-line Boosting实现物体跟踪,据此分析了On-line Boosting算法、Absolute Haar特征、Haar-like特征、弱分类器设计、On-line Boosting跟踪流程。本文对原On-line Boosting算法的重要性权值更新策略进行了改进,并采用上述粒子滤波器加快原On-line Boosting算法的跟踪速度。数据关联部分,考虑到跟踪器-观测值对间的距离、尺度、速度以及运动方向等都对跟踪器-观测值对间的匹配程度会产生影响,对传统的全局最近邻法的匹配函数进行了改进,从而提高了数据关联的准确性和鲁棒性。本文综合以上各个部分,在PC机上搭建了视频监控算法平台,对室内与户外环境进行了大量实验与分析。实验结果表明了上述算法的有效性。更多还原

【Abstract】 Multi-object tracking is a hot research field in video surveillance. Particularly, inrecent years, video surveillance system has played an increasing important role, whichis widely used in homes, car parks, public places, banks and some other places for real-time monitoring. In this paper, we present a detection and tracking integrated videosurveillance system that is used to monitor the outdoor or indoor scenes occasions witha fixed color cameras. This system is able to automatically track varying number oftargets and automatically complete the initialization and termination of the track.Based on the integration of motion detection and object tracking, our video surveil-lance system is divided into four parts: motion detection module, clumps detectionmodule, tracking module and the trajectory generated modules. The main contents ofthis thesis are as follows:Motion detection part: first of all, we analysis the codebook based backgroundsubtraction algorithm in detail. We assume that pixels follow a Gaussian distribu-tion i the time domain, according to statistics and human vision system, and redesignthe codebook, brightness distorition, update rules and etc. based on the original ap-proach. So that it can detec a more complete foreground object. Secondly, we deeplyconstrue Bayesian classification based background subtraction algorithm. Using thecolor-space model in the codebook algorithm, we propose a new thresholding method,which is helpful for removing the moving shadows to some extent.Target tracking part: based on motion detection we propose a framework to ad-dress multi-object tracking problem, that is to decompose multi-object tracking taskinto the combination of multiple single-object tracking tasks. On the one hand, thesingle-object tracking problem could be viewed as a optimal state estimation problem.We study how to utilize particle filter for multi-target tracking in detail. Accordingly, we describe the object tracking related algorithms and implementation. In order toimprove tracking accuracy and efficiency, we propose to use a combination of colorand motion feature as the likelihood function and to use MCMC after particle re-sampling.On the other hand, the single-object tracking problem could also be viewedas a classification problem. Then we study how to use On-line Boosting to achieveobject tracking, and present the analysis and implementation of On-line Boosting al-gorithm, Absolute Haar features, Haar-like features, the weak classifier design, On-line Boosting Tracking processing. Considering the distance, scale, speed and motiondirection of tracker-observation pair would have an impact on the match degree oftracker-observation pair, we propose an improvement on the traditional global near-est neighbor matching function, thereby increase the accuracy and robustness of dataassociation.Based on the above, we build a video surveillance platform on PC, which inte-grates a variety of algorithms and technique, and execute a large number of experi-ments and analysis in indoor and outdoor environment. Experimental results show theeffectiveness of our system.更多还原