CASIA OpenIR  > 毕业生  > 博士学位论文
不同环境下的实时形状跟踪
其他题名Title of Dissertation:Real-time Shape Tracking under Various Environments
曾智洪
学位类型工学博士
导师马颂德
2002-05-01
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业模式识别与智能系统
关键词视觉跟踪 主动轮廓 粒子滤波 Visual Tracking Active Contour Particle Filtering
摘要视觉跟踪是计算机视觉领域中一个非常活跃的研究方向。视觉跟踪一直存在一些难题, 如:由于目标的运动和光照的变化导致在图像序列中目标的形状和表面亮度也处于变化之 中:图像背景中的各种干扰;实时性的要求。在不同的环境下这些问题的难度是不同的。论 文针对不同的环境提出了不同的实时鲁棒的视觉跟踪算法。论文以跟踪物体的形状特征为 手,辅助以物体的表面灰度特征。论文的主要工作有: 1、基于传统的主动轮廓模型(active contour model,Kass et a1.1987),提出了改进的greedy 快速算法(greedy algorithm,Williams and Shan 1992),引入卡尔曼滤波预测,并和区域匹配方法 结合来实现实时的目标跟踪。在区域匹配模块中,采用了自适应匹配密度的策略来控制计算 时间。改进的greedy算法加快了主动轮廓收敛到物体边界的速度。此外,由主动轮廓来指导更 新区域模板,使得算法能够容忍目标的部分遮挡和光照变化。(发表文章1,Zeng and Ma 2000) 2、分析了传统的主动轮廓模型的缺点,通过引入主动轮廓的形状空间(shape space,Blake and Isard 1998)来提高跟踪的稳定性和效率,并在实验中提出了减少遮挡影响的方法。在此基 础上,完成了一个高效的视觉系统来自动跟踪和检测高速公路上同时行驶的多辆汽车。该系 统主要包括四个模块,即车道检测、多个基于可变换两维模型的跟踪器、汽车检测、系统协办 调器。实验采用PETS2001(the Workshop of Performance Evaluation of Tracking and Surveillance 2001)提供的数据验证了系统的鲁棒性和实时性。(发表文章2,Zeng and Ma 2002a) 3、当物体运动在杂乱背景中,跟踪问题就变成了一个非线性、非高斯状态估计问题。在 这种情况下,粒子滤波(particle filtering,Isard and:Blake 1996,Doucet et a1.2001)方法作为一种 基于模拟的方法,提供了估计目标状态后验概率方便而有效的途径。但是,当状态的搜索空 间很大时,例如包含多个特征的联合状态空间,传统的粒子滤波方法所需的粒子数将大大增 加,导致大量的计算时间消耗。为了克服传统粒子滤波器的这一缺点,在分解式图形模型框 架下,我们提出了主动粒子滤波器方法,并应用于实时跟踪图像序列中的目标的多个特征。 在主动粒子滤波器中,每一个粒子通过局部最大值搜索将移动到似然概率局部最大值对应的 状态。这样,比(Wu and Huang 2001)采用的传统粒子滤波器中粒子的使用效率更高,从而导 致所需
其他摘要Visual tracking is one of the hot topics in computer vision. It involves a quite number of diffi- culties, which include: a large variation of a target's shape and appearance in images due to the target movement and illumination change; distraction from environments; real-time demand. This thesis is intended to develop different efficient and robust visual tracking techniques under various environ- ments. The work is mainly concentrated on tracking an object based on its shape attributes, with the help of its appearance ones. The thesis can be summarized as follows: 1. Based on active contour model (Kass et al. 1987), a modified greedy fast algorithm is pre- sented and combined with region correlation to realize real-time tracking. In the region correlation module, an adaptive matching density strategy is adopted to control the computational time. And the modified greedy algorithm speeds up the convergence of an active contour to an object's boundary. Active contour model guides updating the region template so that the tracking method can tolerate partial occlusion and illumination change. Kalman filtering is used to predict the object's movement. (Zeng and Ma 2000) 2. After an analysis on the weakness of traditional active contour mode, shape space of active contour (Blake and Isard 1998) is used to improve the tracker's performance, and a method is pro- posed to reduce the influence of partial occlusion. Then, an efficient vision system is implemented to detect and track multiple cars on the highway. The main modules of the system include: lane detec- tion, separate 2D model-based trackers, heuristic car detection, and a process coordinator. The dataset of PETS2001 is used to testify the robustness and efficiency of the system. (Zeng and Ma 2000a) 3. When an object is moving in cluttered environments, tracking becomes a problem of non- linear and non-Gaussian state estimation. In this situation, as a simulation-based method, particle fil- tering (Isard and Blake 1996; Doucet et al. 2001) provides a convenient and attractive approach for computing the state posterior distribution. However, when the search state space is large, such as one including multiple modalities, the traditional particle filtering needs a large number of particles, con- sequently, the involved computational load becomes unbearably large in practice. In order to deal with the weakness of the traditional particle filtering, an active particle filtering based on the factored graphical model is proposed in this work to track target's multiple modalities in image sequences. In the proposed active particle filtering, the efficiency of every particle is further improved due to the additional local maximum search module, and accordingly the number of required particles is largely reduced in comparison with that of (Wu and Huang 2001). In fact, the number of particle of our active particle filtering is determined mainly by the cluttered
馆藏号XWLW673
其他标识符673
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/5736
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
曾智洪. 不同环境下的实时形状跟踪[D]. 中国科学院自动化研究所. 中国科学院研究生院,2002.
条目包含的文件
条目无相关文件。
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[曾智洪]的文章
百度学术
百度学术中相似的文章
[曾智洪]的文章
必应学术
必应学术中相似的文章
[曾智洪]的文章
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。