【作者】 万琴；

【导师】 王耀南；

【作者基本信息】 湖南大学 ， 控制理论与控制工程， 2006， 硕士

【摘要】 智能视觉监控就是要用计算机视觉的方法,在不需要人为干预的情况下,通过对摄像机拍录的图像序列进行自动分析,实现对动态场景中目标的定位、识别和跟踪,并在此基础上分析和判断目标的行为,从而做到既能完成日常管理又能在异常情况发生的时候及时做出反应。智能视觉监控的一个主要研究内容就是运动分割和跟踪。本论文主要研究了固定场景中多运动目标的检测及跟踪方法。分析了目标检测的常用方法,提出将背景减除与时间差分方法结合的检测目标方法,提高了目标检测准确性。研究了基于色度、梯度建立背景模型检测目标的方法,能实时更新背景模型,基本避免场景中光线变化的影响,有效消除了目标阴影,能准确检测目标,并得到目标轮廓及提取视频目标,为下一阶段针对目标区域的处理提供较完整的目标信息。在目标跟踪阶段,分析了跟踪的基本原理。对目标的线性运动模型,讨论了如何应用Kalman滤波器得到目标状态估计、预测,在目标数据相关(匹配)过程中,采用Mahalanobis距离得到目标运动相似性的度量,提出建立帧间“匹配矩阵”并由匈牙利算法求得对应匹配目标。详细分析了如何将颜色作为目标特征跟踪:采用颜色直方图表示目标颜色特征,由直方图相交算法得到颜色相似性度量。给出了基于目标模型(运动、颜色、形状模型)的跟踪算法。当多目标相互遮挡时,改进了基于目标模型的跟踪算法,在目标运动模型匹配中,由目标运动的概率分布特性,产生多个“虚目标”,确定最佳“虚目标”恢复丢失的遮挡目标,并通过概率方法分割遮挡目标的可见部分,进一步确定遮挡目标的信息。最后,给出了多目标在场景中的不同运动情况及相应跟踪流程。通过对实时视频的实验,证明本论文提出的算法能较好地实现在固定场景中检测多运动目标,并在目标不同运动情况下跟踪各个目标。更多还原

【Abstract】 The aim of Intelligent Visual Surveillance is to make it possible using computer vision method, in the situation of not needing interactive operation, by automatically analyzing the image sequences captured by cameras, computer can localize, track and recognize moving objects in dynamic scenes, analyze and judge the behavior of the objects, thereby not only can complete daily management,but also can react to exception in time. One main research content is motion segmentation and track.This thesis mainly research on the method of detecting and tracking of multiple objects in stationary scenes. It analyzes the common method of detecting objects, and presents a algorithm of detecting objects which combine background subtraction with time difference to improve the performance of detecting. It researches the background model based on gradient and chromaticity which can update background in real-time, which provide exact information of objects to next module.In tracking objects module, the thesis applies Kalman filter to estimate and predict the objects’moving state. In association module, the Mahalanobis distance is utilized to get the similarity among objects.“Matching matrix”is presented to get the best matching object. Then, the thesis analyzes how to use color as a cue to track. The color histogram and intersecting algorithm is applied. When objects are occluded, many patches are produced according to object’s moving probability, the association module based on objects’model define which is the best patch can represents the occluded object. Then the occluded object’s visible parts is segmented by probability. At last, the thesis presents different moving cases and gives corresponding flowchart.According to the experimental results in real-time video, the methods which the thesis presents are proved that it can realize detecting and tracking objects under their different moving cases in stationary scenes.更多还原