中国科学院光电技术研究所机构知识库
Advanced  
IOE OpenIR  > 光电技术研究所博硕士论文  > 学位论文
题名:
复杂背景下运动目标识别和稳定跟踪技术研究
作者: 杨昕梅
学位类别: 博士
答辩日期: 2007-05-31
授予单位: 中国科学院光电技术研究所
授予地点: 光电技术研究所
导师: 吴钦章
关键词: 目标识别 ; 目标跟踪 ; 特征不变量 ; 仿射变换 ; 目标分割 ; 特征点匹配
其他题名: Technologies of Recognizing and Tracking Motion Object under Complex Background
学位专业: 信号与信息处理
中文摘要: 基于图像的目标的识别和跟踪技术已经广泛应用于国防与国民经济建设的诸多领域,融合了计算机视觉、图像处理与模式识别以及计算机应用等相关领域的先进技术和研究成果。对于识别,提取反映目标本质属性和具有较强适应性的特征,一直是研究的核心内容之一。在平移、旋转、尺度等变换下保持不变的常规特征己经不能满足新的目标识别的需要,近年来兴起的3D不变性研究正成为计算机视觉中一个非常活跃的研究领域。在实际的应用中,特征空间往往是非线性可分的,识别的方法要具有较强的分类能力。对复杂背景下运动目标的识别跟踪,尤其是在中远距离观测的时候,由于许多不可预知的环境条件(光照、气候、能见度等)、背景噪声干扰等的影响,影响了目标跟踪的稳定性。因此如何实现三维运动目标的识别,以及在复杂背景下的稳定跟踪,成为本文的研究重点。 在简要介绍运动目标识别与跟踪技术的发展概况、相关技术的基本理论和方法的基础上,对运动目标特征不变量、目标识别方法、以及几种保证稳定的跟踪方法等方面进行了探讨和研究。 在提取运动目标特征不变量研究方面,首先分析了几种二维平面运动目标不变量。之后,着重研究了3D不变特征。通过分析了目标成像的模型及仿射与透视变换的关系,提出了具有仿射不变性的离散余弦描述子。这种方法使用微分面积对目标轮廓进行参数化,进而计算其离散余弦变换。最后,所得到的一系列描述子具有仿射不变性,尤其适用于在视点发生变化时,三维空间中运动目标识别。通过理论分析,给出了算法的推导过程,并对三维空间中不同姿态的飞机轮廓图像进行了仿真实验。实验结果表明,各描述子描述的特征值是相对稳定的。 分类器的设计也是目标识别的关键之一。采用基于核的Fisher分类方法,实现了非线性可分。在这种方法中使用了支持向量机中的核函数思想,利用核函数把输入的样本映射到某一高维空间,然后在此特征空间进行Fisher线性判别。提出将余弦描述子仿射不变量作为目标特征量,结合基于核的Fisher分类方法对三维运动目标进行识别。通过实验表明,这种方法有效的提高了识别率。 除了用特征分类器的方法识别目标外,通过相关性也能很好的实现目标的识别。提出用仿射不变轮廓线作为特征进行识别。首先目标轮廓曲线用具有仿射不变性的参数将其参数化。然后,待识别目标的参数化轮廓线正交投影后,可使其与库存同一型号的目标轮廓线保持较强的相关性。最后,用归一化相关函数作为识别准则确定目标类别。对于空间姿态变化前后的同一类目标,它们的相关度大于0.9。 在运动目标跟踪中,对于尺寸较小的目标利用质心跟踪可以获得良好的效果。针对复杂背景下的质心跟踪问题,提出了一种目标分割的算法。对相邻两帧图像的跟踪框和背景框的灰度分布统计,加入滤波因子, 再根据Bayes判别规则构造判别函数,把背景像素和目标像素进行分类。实验结果证明,所提出的方法克服了由于图像序列中某一帧图像灰度分布发生突变产生的影响,能较好的分割出图像中的目标。 在多模复合图像跟踪中,要采用多种跟踪方法保证准确的跟踪,因此,在目标质心跟踪不稳定时,采用基于特征点匹配的跟踪方法。采用SUSAN方法检测目标边缘,提出了基于多特征角点检测提取角点,提高检测精度和抗噪性。在不同情况下采用不同的特征点,用改进的Hausdorff距离进行匹配跟踪。改进了跟踪策略,保证了在复杂背景下,即使出现严重的遮挡问题也能稳定的跟踪。 针对复杂背景下目标的识别与稳定跟踪的难点,结合工程实际,提出了针对性很强的算法和方案。在光电探测和图像处理领域具有重要的现实意义。
英文摘要: Object recognition and tracking technique based on images has been widely used in many fields of defense and civil economic constructions. It combines advanced technologies and research achievements in computer vision, image processing, pattern recognition, compute application and other relative fields. Extracting features that show object's intrinsic characteristics and have strong adaptabilities is the kernel of recognition research all along. The traditional invariant features in translation, rotation and scaling transformation can't satisfy the needs of advanced object recognition. So the research about 3D invariance prospered in recent years is becoming one of the active research fields in computer vision. In practical application, the feature space is generally non-linear divisible. So the method of recognition need have strong classification ability. It affects the performance of stable tracking in complex environment that because of the unpredictable circumstance (light、weather、visibility) and noise interference , especially from long distance. So it become the emphases in this research that how to realize recognizing motion objects in 3D space and stable tracing under the complex background. Based on the brief introduction of the development of motion object recognition and tracing, and its relative fundamental theories and method, the discussion and research is processed from several aspects: feature invariant of object, the method of object recognition and stable tracing and so on. In the beginning of the research of extracting feature invariant of motion object, several invariant in 2D space were analyzed. Then the research was emphasized on 3D invariant. Through the analysis of the imaging model and the relation between affine and perspective transformation, Affine-Invariant Discrete Cosine Transformation Descriptor was proposed. Firstly, the boundary of an object was parameterized by differential area, and then the parameterized boundary was transformed by discrete cosine transformation. Finally, based on this method, a set of descriptors was obtained, which have affine invariant. Especially, they are invariant as the change of viewport in recognition of 3D motion object. Through analysis of theory, the computational process of this method was given. Many planes of different poses in 3D space were simulated. The experiment results show that values of each kind of character are stable correspondingly. The design of classifier is also one key problem of object recognition. Using the method based on kernel Fisher discriminant can realize non-linear classification. The method used the same thinking of kernel function as in Support Vector Machine. After mapping the data non-linearly into a certain higher dimensional feature space, Fisher’s linear discriminant was used in this feature space. A method of 3D motion object recognition was proposed. Affine-Invariant Discrete Cosine Transformation Descriptor was taken as feature, combining kernel Fisher discriminant. The experiment results show that the method improves the discrimination effectively. Except using feature classifier to recognize object, the method based on correlation is a better way to realize object recognition. A method of regarding contour with affine invariant as feature was proposed to recognizing. Firstly, contour curve of object was parameterized by an affine variable. Then parameterized curve of object waiting for recognition was projected orthogonally. So it could keep better correlation with the object of the same class in store. Finally, normalized related function was regarded as the criterion of recognition. The experiment results show that the correlation of objects of the same class is more than 0.9 in the 3D’s change of pose. In the motion object tracing, centroid tracking is better to the object of smaller size. In order to realize centroid tracking in real-time images,a method of segmentation is proposed. It combined the statistic of the gray scale in track and background gate of two contiguous frames with a smoothing value. Furthermore, according to the classification criterion built by Bayesian decision theory, the background and the object were classified. The experiment results show that this proposed method can avoid the effect that attributes to the wave of gray scale in some frames of an image sequence, and can segment the object from an image better. Multi-mode tracking techniques fuse multiple tracking algorithms in order to promise exact tracking. So, when centroid tracking was not stable, the method based on feature points matching was used. Edge was detected by SUSAN principle. A new method of corner detect basing multi feature was proposed. It has higher precision and strong noise rejection. The different feature points were used in different conditions, and combining the improved Hausdorff distance to matching. A project of tracing object was improved. It promises stability of tracking in the complex background even in the condition of appearing the problem of occlusion. In conclusion, some innovated methods are developed to solve the difficulties in the object recognition and tracking and engineering practicality. The work of this dissertation has important reference to the future development of optics-electronics detection and image processing.
语种: 中文
内容类型: 学位论文
URI标识: http://ir.ioe.ac.cn/handle/181551/231
Appears in Collections:光电技术研究所博硕士论文_学位论文

Files in This Item:
File Name/ File Size Content Type Version Access License
10001_200418015127027杨昕梅_paper.doc(16173KB)----限制开放View 联系获取全文

Recommended Citation:
杨昕梅. 复杂背景下运动目标识别和稳定跟踪技术研究[D]. 光电技术研究所. 中国科学院光电技术研究所. 2007.
Service
Recommend this item
Sava as my favorate item
Show this item's statistics
Export Endnote File
Google Scholar
Similar articles in Google Scholar
[杨昕梅]'s Articles
CSDL cross search
Similar articles in CSDL Cross Search
[杨昕梅]‘s Articles
Related Copyright Policies
Null
Social Bookmarking
Add to CiteULike Add to Connotea Add to Del.icio.us Add to Digg Add to Reddit
文件名: 10001_200418015127027杨昕梅_paper.doc
格式: Microsoft Word
此文件暂不支持浏览
所有评论 (0)
暂无评论
 
评注功能仅针对注册用户开放,请您登录
您对该条目有什么异议,请填写以下表单,管理员会尽快联系您。
内 容:
Email:  *
单位:
验证码:   刷新
您在IR的使用过程中有什么好的想法或者建议可以反馈给我们。
标 题:
 *
内 容:
Email:  *
验证码:   刷新

Items in IR are protected by copyright, with all rights reserved, unless otherwise indicated.

 

 

Valid XHTML 1.0!
Copyright © 2007-2016  中国科学院光电技术研究所 - Feedback
Powered by CSpace