2-D Feature-Point Selection and Tracking Using 3-D Physics-Based Deformable Surfaces

876IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY,VOL.17,NO.7,JULY2007 2-D Feature-Point Selection and Tracking Using3-D Physics-Based Deformable Surfaces

Michail Krinidis,Nikos Nikolaidis,and Ioannis Pitas

Abstract—This paper presents a novel approach for selecting and tracking feature points in video sequences.In this approach, the image intensity is represented by a3-D deformable surface model.The proposed approach relies on selecting and tracking feature points by exploiting the so-called generalized displacement vector that appears in the explicit surface deformation governing equations.This vector is proven to be a combination of the output of various line-and edge-detection masks,thus leading to distinct, robust features.The proposed method was compared,in terms of tracking accuracy and robustness,with a well-known tracking al-gorithm,Kanade–Lucas–Tomasi(KLT),and a tracking algorithm based on scale-invariant feature transform(SIFT)features.The proposed method was experimentally shown to be more precise and robust than both KLT and SIFT tracking.Moreover,the fea-ture-point selection scheme was tested against the SIFT and Harris feature points,and it was demonstrated to provide superior results. Index Terms—Feature point selection,tracking,intensity sur-face,video analysis,3-D deformable models.

I.I NTRODUCTION

T RACKING objects in video sequences is a frequently encountered task in video-based applications,such as surveillance,hand-gesture recognition,human–computer inter-action,smart environments,motion capture for virtual reality and computer animation,video editing,medical and meteorological imaging,and3-D scene reconstruction from uncalibrated video. Thus,in the last two decades,intensive research has been carried out in this area.Building a tracking system is far from being a simple process due to varying lighting conditions,partial occlu-sions,clutter,unconstrained motion,and so on.So far,various systems for person,face,and object tracking have been presented in the literature.These systems can be broadly divided in four categories:

?color-based tracking;

?template-based tracking;

?contour tracking;

?feature-based tracking.

Additional information about the aforementioned tracking cat-egories can be found in the excellent review publications that have appeared in the literature[1]–[5].

Color is a distinctive object feature and,therefore,is useful for object localization on static and video images.Color in-

Manuscript received December19,2005;revised May18,2006,September 18,2006,and January4,2007.This work was conducted in conjunction with the“SIMILAR”European Network of Excellence on Multimodal Interfaces of the IST Programme of the European Union.This paper was recommended by Associate Editor P.L.Correia.

The authors are with the Department of Informatics,Aristotle University of Thessaloniki,54124Thessaloniki,Greece(e-mail:mkrinidi@aiia.csd.auth.gr; nikolaid@aiia.csd.auth.gr;pitas@aiia.csd.auth.gr).

Digital Object Identi?er10.1109/TCSVT.2007.897463formation produces satisfactory tracking results and allows fast processing,which is important for a tracking system that needs to run at a reasonable frame rate.Many approaches are based on color histograms while some others use global color reference models[6],[7].

Template matching techniques are used by many researchers to perform object tracking by following the same principles with the template matching techniques used in object recognition [8],[9].Template-based tracking involves the use of multiple templates or template warping to accommodate changes in ob-ject pose.The process of determining correspondences between image and template pixels is computationally expensive but pro-vides robust tracking results.

Tracking using outline contour information is easier than modeling and tracking the entire object area,e.g.,when using color.Moreover,contour tracking is more robust than using simple corner or edge tracking,since it can be adapted to cope with partial occlusions.The active contour representation introduced by Kass et al.[10]is the most popular method for contour delineation and tracking.

Feature-based tracking is a frequently used approach,in which moving objects are represented by feature points de-tected prior to tracking or during tracking.Feature-based tracking,though prone to tracking errors,can be implemented very ef?ciently and is important in many time-critical appli-cations.The selection of the feature points depends on the algorithm and usually is based on speci?c features(local image properties)of these points.A signi?cant number of feature tracking algorithms have been introduced trying,in general, to correlate image features from frame to frame.In[11],the feature points are stochastically selected based on the energy of their Gabor wavelet transform coef?cients.The global place-ment of the feature points is determined by a2-D mesh,using the area of the triangles formed by the feature points.This method uses a local feature vector containing Gabor wavelet transform coef?cients and a global feature vector containing triangle areas.In order to?nd the corresponding features in the next frame,the2-D golden section algorithm is employed. Mid-level features(strokes)are used in[12]instead of low-level ones(edge points).Strokes are accomplished by organizing edge points through an edge-linking operation.Two labels (valid/invalid)are considered for each stroke,and a probability is assigned to each of them.In this way,all of the strokes con-tribute to track the moving object but with different weights.In [13],multiple features were used in order to improve the ac-curacy and robustness of a real-time tracker.More speci?cally, color histogram features combined with edge-gradient-based shape features were tracked over time under a Monte Carlo

1051-8215/$25.00?2007IEEE

KRINIDIS et al.:2-D FEATURE-POINT SELECTION AND TRACKING USING3-D PHYSICS-BASED DEFORMABLE SURFACES877

framework.A comparison of four feature-point tracking al-gorithms is given in[14].Many researchers,instead of trying to improve the tracking performance through the selection of “good”features,exploited the knowledge of how a tracker works and tried to impose several constraints so as to improve the tracking of feature points[15]–[17].

In most of the cases,an initialization step which depends on the tracking algorithm is applied prior to tracking and de?nes the area of points that will be tracked.In feature-based algo-rithms,several feature-point selection strategies can be used. The goal is to obtain distinctive feature points on the image that are appropriate for tracking.Many of these feature points are also used for image matching applications,e.g.,for?nding the correspondences between two views of the same scene.Lowe proposed the scale-invariant feature transform(SIFT)feature points[18],which are scale-,rotation-,and partially illumi-nation-invariant.The SIFT feature points were used for image matching and image retrieval.Harris et al.[19]proposed a com-bined edge and corner detector which provides feature points that exhibit high“cornerness”and thus are suitable for tracking. In[20],the feature points are extracted based on the eigen-values of an image gradient matrix constructed over a window around the candidate feature point.If the minimum eigenvalue of this matrix is larger than a user-de?ned threshold,then the feature point is considered to be good for tracking.The ex-tracted feature points are optimal for the tracking algorithm pre-sented in the same paper.Moreover,a scheme for the selec-tion of discriminative tracking features was proposed in[21]. Given a set of features,the log-likelihood ratios of class condi-tional sample densities from the objects of interest and the back-ground were computed,to form a new set of candidate features tailored to the local object/background discrimination task.The two-class variance ratio is used to rank these new features ac-cording to how well they separate sample distributions of object and background pixels.This feature evaluation mechanism is embedded in a mean-shift tracking system that adaptively se-lects the top-ranked discriminative features for tracking.

A novel feature selection and tracking algorithm is proposed in this paper.The approach was motivated by the technique pre-sented in[22]–[24],which aims at analyzing nonrigid object motion,with application to medical images.Nastar et al.[22] used deformable models to approximate the dynamic object sur-face deformations in time sequences of volume data(i.e.,se-quences of3-D data)and applied modal analysis techniques(a standard engineering technique that allows more effective com-putations and provides a closed-form solution of the deforma-tion process)in order to describe and analyze the deformations. The framework proposed in this paper has been also exploited for the alignment of serially acquired slices[24],for multi-modal brain image analysis[23]and segmentation of2-D ob-jects[25].In our case,the deformable model formulation is used in a totally different and novel application,i.e.,that of feature-point tracking.We assume that the image intensity in each video frame can be approximated by a deformable“intensity”sur-face,where we select and track characteristic feature points.The proposed technique exploits a byproduct of the explicit surface deformation governing equations,in order to select and subse-quently track distinctive feature points.More speci?cally,the feature-point selection process utilizes the so-called generalized displacement vector[22],which is shown to be a novel combi-nation of the output of various line and edge detection masks and,thus,produces feature points corresponding to local edges, lines,corners,or other characteristic image features that are suit-able for tracking.The connection between the deformable sur-face model and the line/edge detector operators is an important outcome of this study.The tracking procedure that follows the feature selection is based on measuring and matching the gen-eralized displacement vector of the feature points from frame to frame.

In summary,the novelty of this paper lies in the use of a de-formable surface to approximate the image intensity surface and the consequent use of a term appearing in the deformation pro-cedure to perform robust feature-point selection and tracking. With respect to the deformable model and its modal analysis as introduced in[26]and[22]and further used in[23]and[24], the novelty lies in the use of the model for a different applica-tion(i.e.,that of feature selection and tracking),the use of dif-ferent external forces that attract the model towards the image intensity(as will be described in Section II)and the use of an intermediate result(generalized displacement vector)of the de-formation procedure instead of using the model per se. Compared with existing feature selection and tracking algo-rithms,namely the Kanade–Lucas–Tomasi(KLT)[27]and the SIFT algorithm[28],the proposed method achieves better per-formance in terms of tracking accuracy and robustness.The re-sults show that the proposed method is robust against rotations, zooming,varying lighting conditions,and hard shadows and can track the selected features for long time periods.Moreover,the feature-point selection part of the algorithm was compared with other feature selection algorithms,i.e.,the SIFT[18]and Harris [19]feature-point detectors,and was shown to provide superior results in their subsequent tracking.

The remainder of this paper is organized as follows.In Section II,a brief description of the deformation procedure based on modal analysis is presented.The feature-point se-lection procedure is introduced in Section III.The tracking algorithm is described in Section IV.The performance of the proposed technique,as well as a comparison between the proposed algorithm and the well-known KLT feature-based tracking algorithm[27],tracking using SIFT feature points [18],and feature selection using SIFT and Harris[19]feature detectors are presented in Section V.Final conclusions are drawn in Section VI.

II.D EFORMABLE M ODEL D ESCRIPTION

Image

intensity can be assumed to de?ne a surface over the image

domain that will be subsequently called intensity surface.The proposed tracking approach focuses on parameterizing the3-D space de?ned

by that is called

the space[29].A3-D physics-based deformable surface model,introduced in[22],[23],and[26],is used for this purpose.In this section,the methodology described in these papers will be brie?y reviewed,so as to make this paper self-contained.For more details,including the assumptions that are involved,interested readers can consult the abovementioned papers.

微生物学课件

微生物学第一章绪论介绍以下几个内容：系统与进化生物学的三域理论、微生物的定义、微生物学分支、微生物学的建立和发展、微生物与生物技术、学习本课程的主要参考书系统与进化生物学三域学说或三域理论进化生物学：研究生命起源与进化关系的生物学重要分支。生物系统发育：探究自然界生物在性状上的相似程度，依据相似程度建立一定的系统发育关系(聚类,Cluster）。以树形结构表示系统发育关系或进化关系（phylogeny)。系统发育树（Phylogenetic tree）又称为进化树，以此树表示生物的亲缘进化关系，每个节点代表其各分支的最近共同祖先，而节点间的线段长度对应进化距离或估计进化时间。系统树有有根树和无根树。 三领域理论:1977年，Woese 和Fox根据核糖体 RNA(rRNA)小亚基（SSU) 序列比对结果，构建了生物 界的三域系统树，根据系统 树提出將生物分为三个域： 细菌域(Domain Bacteria)、古 菌域(Domain Archaea)和真 核生物域(Domain Eukarya)， 此即为三域系统 (Three-domain system)。问 题：为什么Woese和Fox要 选择核糖体RNA(rRNA)小 亚基（SSU)序列？ 微生物定义及微生物学微生物（Micro organism，microbial）——通常是指不借助显 微镜用肉眼看不见的个体微小、结构简单的微 小生物，包括病毒（Virus)、细菌（Bacteria)、 真菌(Fungi)、原生动物(Protists)、单细胞藻类 (Unicell algae )。 病毒：非细胞生物，没有新陈代谢，非完整生 命。原核微生物：细菌、古菌，细胞生物， 但没有细胞核和细胞器。真核微生物：真菌， 单细胞藻类，原生动物，细胞生物，有细胞核 和细胞器。微生物的度量单位：有细胞结构 的为微米（um)，病毒为纳米（nm)。 微生物共同特点：（小、多、快、旺、强、广） 1）体积小：细胞微生物的细胞大小一般以微米（μ m)、非细胞微生物病毒大小以纳米(nm)来计算； （2）吸收多，代谢旺：生理代谢迅速，吸收的物 质转化速度快。 （3）生长旺，繁殖快如大肠杆菌 （4）易变异，如感冒病毒 （5）适应强，快速适应环境的变化。 （6）分布广，种类多。（具估计，细菌有100多万 种，真菌有１００多万种，因而，为生物多样性研 究和资源开发提供极丰富的材料）。 微生物学（Microbiology）：是研究微生物生命活动规律的学科,是生物科学的一个极为重要的分支。微生物学：讲解微生物学的一些基础理论和基本实验（试验）

医学微生物ppt医学微生物学课件

医学微生物ppt医学微生物学课件 医学微生物资料 该资料主要是以该科教师课件里的内容为主，加以和，能力有限，不免会有不少错误，望大家能理解。 绪论章节 目的要求： 1. 掌握微生物的定义和分类; 2. 了解微生物学的发展史及奠基人的工作; 3. 熟悉医学微生物学的定义. 本篇主要结构：一、微生物的分类 非细胞型微生物原核细胞型微生物真核细胞型微生物 二、微生物的发展史（主要人物） 1) 荷兰列文虎克：发现微生物

2) 法国巴斯德：发现微生物与人、动植物疾病的关系 3) 德国科赫：创立细菌分离培养和染色法，并发现多种致病菌4) 弗莱明：发现青霉素 医细学真病菌微毒学生B物MVa学yi cr to 2 微生物的分类 一、 非细胞型微生物:无细胞结构，无产生能量的酶系统，含一种核酸(DNA或 RNA); 只能在活细胞内增殖; 是最小一类微生物. (病毒是典型代表，还有比病毒更简单的类病毒和朊粒)。

非细胞型微生物 二、原核细胞型微生物:核呈环状裸DNA团块结构，无核膜、核仁。细胞器很不 完善, 只有核糖体。DNA 和 RNA同时存在。原核细胞型微生物的种类包括细菌、支原体、衣原体、立克次体、螺旋体和放线菌等。 三、 真核细胞型微生物:细胞分化程度高，有核膜和核仁结构，进行有丝分裂，胞 质内细胞器完整，同时含有两类核酸。真菌属此类。 课后思考题 1) 定义：微生物、医学微生物学 2) 微生物分哪3类，有什么区别？并举例说明。 3) 微生物的主要两位奠基人是谁？

4) 近年医学微生物学得到了迅速地发展，主要表现在哪些方面？ 5) 医学微生物学未来发展方向是什么？ 3 第1章细菌的形态与结构 目的要求： 1.掌握细菌 L 型、质粒、荚膜、鞭毛、菌毛、芽胞、异染颗粒、外膜蛋白的定义 2. 掌握细菌的测量单位及细菌的基本形态和结构 3. 掌握革兰阳性菌与革兰阴性菌细胞壁结构的区别 4. 掌握细菌特殊结构的医学意义 5. 了解革兰染色原理，掌握革兰染色的步骤、结果、医学意义 6. 熟悉细胞膜、细胞质、核质的结构、功能 本篇主要结构： 第一节细菌的大小与形态