CORC  > 自动化研究所  > 中国科学院自动化研究所  > 毕业生  > 博士学位论文
题名视觉监控中的目标跟踪研究
作者李敏
学位类别工学博士
答辩日期2010-11-28
授予单位中国科学院研究生院
授予地点中国科学院自动化研究所
导师谭铁牛
关键词Omega形状 方向滤波的金字塔统计量 简化的生物启发特征 姿态估计器 增量自调节粒子滤波 底层启示模型 MCMC粒子滤波 Omega-shape Pyramidal Statistics of Oriented Filtering Simplified Biologically Inspired Features Pose Estimator Incremental Self-tuning Particle Filtering low-level cues MCMC-particle filtering
其他题名Object Tracking for Visual Surveillance
学位专业模式识别与智能系统
中文摘要智能视觉监控作为安全监控的一个有效手段,越来越受到各国政府的重视。其终极目标是让计算机成为人的大脑,让摄像头成为人的眼睛,由计算机智能地分析摄像头所获取的图像序列,对场景内容进行理解,实现对异常行为的自动报警和预警。智能视觉监控涉及计算机视觉的底层和高层部分。底层视觉主要是对场景中感兴趣的目标进行检测、跟踪和识别,而高层视觉则在底层视觉的基础上,对感兴趣目标作行为分析和理解。视觉跟踪属于中底层视觉部分,其任务是在连续的图像序列中对运动目标进行定位,获取目标轨迹和运动参数等信息,是智能监控的关键技术之一。本文对视觉监控场景中的目标跟踪问题进行了广泛的研究,内容涉及目标检测、表观模型、跟踪策略以及多目标跟踪等方面。在本文中,主要工作和贡献有: (1)提出了一个基于头肩Omega形状特征的行人检测方法。该方法采用一个基于Haar特征的级联AdaBoost分类器作粗分类,快速地排除大量的负样本窗口,然后再利用一个分类性能更好的基于HOG(Histograms of Oriented Gradients)特征的AdaBoost分类器对剩下的稀疏窗口作精分类,能快速而准确地检测出监控场景中的人的头肩部分。由于头肩所组成的Omega特征在各种视角下具有一定的不变性,能适应多种场景,具有广泛的应用前景。 (2)提出了一个新的特征集—方向滤波的金字塔统计量(PSOF,Pyramidal Statistics of Oriented Filtering)用于行人检测。虽然HOG(梯度方向直方图)特征对于正常成像条件下的行人检测非常有效,但是实际应用表明,在图像模糊或图像噪声存在的情况下,HOG特征的检测性能会大幅下降。因此,本文设计了一个PSOF特 征集用于目标的特征表达。PSOF特征集避免了使用梯度来表征图像的局部方向信息,而是使用一个Gabor滤波器组来提取方向特征,对噪声有一定的抑制能力;另外,与HOG特征只提取单个尺度上的特征不同,PSOF特征集在图像的多个尺度上提取方向信息(尺度越高,模糊或噪声的影响越小),提高了特征集的鲁棒性。实验表明,在正常的成像条件下,PSOF特征集的检测性能跟HOG特征集相当;而在图像噪声或者图像模糊存在的情况下,PSOF特征集的检测性能要远远好于HOG特征集。 (3)提出了一个基于简化的生物启发特征(SBIF,Simplified Biologically Inspired Features)的表观模型用于目标跟踪。该表观模型使用轻微的位置-/尺度-无关的生物视觉感知特征作为目标表达,采用Bhattacharyya系数来测量目标模型和候选目标之间的相似性,并被结合到一个SIR(Sampling Importance Resampling)粒子滤波框架下,用于基于贝叶斯状态推断目标跟踪。大量的实验表明,SBIF方法对于光照变化、目标姿态变化以及部分遮挡具有较好的鲁棒性;而且也具有跟当前主流的表观模型相当的跟踪精度。 (4)提出了一个仿射群上的增量自调节粒子滤波(ISPF,Incremental Self-tuning Particle Filtering)过程用于目标跟踪。ISPF方法使用类SIFT(Scale Invariant Feature Transform) 描述子作为基本特征,通过增量PCA(Incremental Principal Component Analysis)在线学习一个自适应的子空间作为表观...
英文摘要As an effective means for security surveillance, intelligent visual surveillance (IVS) becomes increasingly popular in the field of computer vision and attracts more and more attentions from the public. The ultimate goal of IVS is to make camera become human’s ”eyes”, and make computer become human’s ”brain”, then construct a vision system that can automatically understand the semantic contents in the observed scene and realize automatic warnings/alarms against abnormal activities. IVS contains both low-level vision process that includes object detection, tracking and recognition, and high-level vision process that usually means behavior analysis and understanding. Visual tracking belongs to the low/intermediate vision process, the task of which is to position object in consecutive image sequence, and to obtain object’trajectory and motion parameters. This thesis studies four basic issues involved in visual tracking: object detection, appearance model, tracking strategy and multi-target tracking. The main contributions include:(1)This thesis proposed an Omega-shape feature based human detection method. In this method, a Haar-feature based cascade AdaBoost classifier is used to rapidly exclude the obvious non-head-shoulder samples, then the remaining sparse samples are classified by a HOG (Histograms of Oriented Gradients) based AdaBoost classifier with much better classification performance. This decision-level fusion can detect the head-shoulder parts of human beings in surveillance scenes very rapidly and accurately. Since the Omega-like shape formed by the head-shoulder is invariant in almost all view angles, the proposed Omega-shape based human detection method can adapt to various indoor and outdoor scenes, and thus has a wide range of potential applications.(2)This thesis proposed a new feature set—PSOF (Pyramid Statistics of Oriented Filtering) for human detection. Though HOG feature has been proved to be a great success for human detection under normal imaging conditions, some studies show that its performance decrease sharply when used in blurred/noised images. Instead of using gradients to represent local orientation information, the PSOF descriptor uses a Gabor filter bank to extract orientation information, which shows some robustness to image noise; in addition, different from one-scale HOG feature, PSOF descriptor uses pyramidal statistics of multi-scale orientation information to represent object’s local shape, which show gr...
语种中文
其他标识符200718014628046
内容类型学位论文
源URL[http://ir.ia.ac.cn/handle/173211/6309]  
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
李敏. 视觉监控中的目标跟踪研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2010.
个性服务
查看访问统计
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。


©版权所有 ©2017 CSpace - Powered by CSpace