外文翻譯---運動圖像和運動矢量檢測綜述

1、<p><b>　　外文文獻：</b></p><p>　　A SURVEY ON MOTION IMAGE </p><p>　　AND THE SEARCH OF MOTION VECTOR</p><p>　　After motion detection, surveillance systems generally track

2、 moving objects from one frame to another in an image sequence. The tracking algorithms usually have considerable intersection with motion detection during processing. Tracking over time typically involves matching objec

3、ts in consecutive frames using features such as points, lines or blobs. Useful mathematical tools for tracking include the Kalman filter, the Condensation algorithm, the dynamic Bayesian network, the geodesic method, etc

4、</p><p>　　A. Region-Based Tracking</p><p>　　Region-based tracking algorithms track objects according to variations of the image regions corresponding to the moving objects. For these algorithms,

5、 the background image is maintained dynamically, and motion regions are usually detected by subtracting the background from the current image. Wren et al. explore the use of small blob features to track a single human in

6、 an indoor environment. In their work, a human body is considered as a combination of some blobs respectively representing various</p><p>　　Laboratory (JPL), and the PATH system developed by the Berkeley gro

7、up.</p><p>　　Although they work well in scenes containing only a few objects (such as highways), region-based tracking algorithms cannot reliably handle occlusion between objects. Furthermore, as these algor

8、ithms only obtain the tracking results at the region level and are essentially procedures for motion detection, the outline or 3-D pose of objects cannot be acquired. (The 3-D pose of an object consists of the position a

9、nd orientation of the object).Accordingly, these algorithms cannot satisfy the require</p><p>　　B. Active Contour-Based Tracking</p><p>　　Active contour-based tracking algorithms track objects b

10、y representing their outlines as bounding contours and updating these contours dynamically in successive frames. These algorithms aim at directly extracting shapes of subjects and provide more effective descriptions of o

11、bjects than region-based algorithms. Paragios et al. detect and track multiple moving objects in image sequences using a geodesic active contour objective function and a level set formulation scheme. Peterfreund explore

12、s a </p><p>　　C. Feature-Based Tracking</p><p>　　Feature-based tracking algorithms perform recognition and tracking of objects by extracting elements, clustering them into higher level features

13、and then matching the features between images. Feature-based tracking algorithms can further be classified into three subcategories according to the nature of selected features: global feature-based algorithms, local fea

14、ture-based algorithms, and dependence-graph-based algorithms.</p><p>　　? The features used in global feature-based algorithms include centroids, perimeters, areas, some orders of quadratures and colors, etc.

15、 Polana et al. provide a good example of global feature-based tracking. A person is bounded with a rectangular box whose centroid is selected as the feature for tracking. Even when occlusion happens between two persons d

16、uring tracking, as long as the velocity of the centroids can be distinguished effectively, tracking is still successful.</p><p>　　? The features used in local feature-based algorithms include line segments,

17、curve segments, and corner vertices, etc.</p><p>　　? The features used in dependence-graph-based algorithms include a variety of distances and geometric relations between features.</p><p>　　The

18、above three methods can be combined .In there cent work of Jang et al. [34], an active template that characterizes regional and structural features of an object is built dynamically based on the information of shape, tex

19、ture, color, and edge features of the region. Using motion estimation based on a Kalman filter,the tracking of a nonrigid moving object is successfully performed by minimizing a feature energy function during the matchin

20、g process.</p><p>　　In general, as they operate on 2-D image planes, feature-based tracking algorithms can adapt successfully and rapidly to allow real-time processing and tracking of multiple objects which

21、are required in heavy thruway scenes, etc. However, dependence-graph-based algorithms cannot be used in real-time tracking because they need time-consuming searching and matching of graphs. Feature-based tracking algorit

22、hms can handle partial occlusion by using information on object motion, local features and de</p><p>　　? The recognition rate of objects based on 2-D image features is low, because of the nonlinear distortio

23、n during perspective projection and the image variations with the viewpoint’s movement.</p><p>　　? These algorithms are generally unable to recover 3-D pose of objects.</p><p>　　? The stability

24、of dealing effectively with occlusion, overlapping and interference of unrelated structures is generally poor.</p><p>　　D. Model-Based Tracking</p><p>　　Model-based tracking algorithms track obj

25、ects by matching projected object models, produced with prior knowledge, to image data. The models are usually constructed off-line with manual measurement, CAD tools or computer vision techniques. As model-based rigid o

26、bject tracking and model-based no rigid object tracking are quite different, we review separately model-based human body tracking (no rigid object tracking) and model-based vehicle tracking (rigid object tracking).</p

27、><p>　　1.Model-Based Human Body Tracking:</p><p>　　The general approach for model-based human body tracing is known as analysis-by-synthesis, and it is used in a predict-match-update style. Firstly

28、, the pose of the model for the next frame is predicted according to prior knowledge and tracking history. Then, the predicted model is synthesized and projected into the image plane for comparison with the image data. A

29、 specific pose evaluation function is needed to measure the similarity between the projected model and the image data. According to di</p><p>　　? Construction of human body models；</p><p>　　? Re

30、presentation of prior knowledge of motion models and motion constraints；</p><p>　　? Prediction and search strategies. Previous work on these three issues is briefly and respectively reviewed as follows.</

31、p><p>　　A．Human body models: Construction of human body models is the base of model-based human body tracking. Generally, the more complex a human body model, the more accurate the tracking results, but the mor

32、e expensive the computation. Traditionally, the geometric structure of human body can be represented in the following four styles.</p><p>　　? Stick figure. The essence of human motion is typically contained

33、in the movements of the torso, the head and the four limbs, so the stick-figure method is to represent the parts of a human body as sticks and link the sticks with joints. Karaulova et al. use a stick figure representati

34、on to build a novel hierarchical model of human dynamics encoded using hidden Markov models (HMMs), and realize view-independent tracking of a human body in monocular image sequences.</p><p>　　? 2-D contour.

35、 This kind of human body model is directly relevant to human body projections in an image plane. The human body segments are modeled by 2-D ribbons or blobs. For instance, Ju et al. propose a cardboard human body model,

36、in which the human limbs are represented by a set of jointed planar ribbons. The parameterized image motion of these patches is constrained to enforce the articulated movement of human limbs. Niyogi et al. use the spatia

37、l-temporal pattern in XYT space to track, anal</p><p>　　? Volumetric models. The main disadvantage of 2-D models is that they require restrictions on the viewing angle. To overcome this disadvantage, many re

38、searchers use 3-D volumetric models such as elliptical cylinders, cones, spheres, super-quadrics,etc. Volumetric models require more parameters than image-based models and lead to more expensive computation during the ma

39、tching process. Rohr [28] makes use of fourteen elliptical cylinders to model a human body in 3-D volumes. Wachter et al. establi</p><p>　　? Hierarchical model. Plankers et al. present a hierarchical human m

40、odel for achieving more accurate results. It includes four levels: skeleton, ellipsoid meatballs simulating tissues and fats, polygonal surface representing skin, and shaded rendering.</p><p>　　B. Motion mod

41、els: Motion models of human limbs and joints are widely used in tracking. They are effective because the movements of the limbs are strongly constrained. These motion models serve as prior knowledge to predict motion par

42、ameters, to interpret and recognize human behaviors,or to constrain the estimation of low-level image measurements. For instance, Bregler decomposes a human behavior into multiple abstractions, and represents the high-l

43、evel abstraction by HMMs built from phases of s</p><p>　　C. Search strategies: Pose estimation in a high-dimensional body configuration space is intrinsically difficult, so, search strategies are often caref

44、ully designed to reduce the solution space. Generally, there are four main classes of search strategies: dynamics, Taylor models, Kalman filtering, and stochastic sampling. Dynamical strategies use physical forces applie

45、d to each rigid part of the 3-D model of the tracked object. These forces, as heuristic information, guide the minimization of the</p><p>　　2. Model-Based Vehicle Tracking: As to model-based vehicle tracking

46、, 3-D wire-frame vehicle models are mainly used. The research groups at the University of Reading, the National Laboratory of Pattern Recognition (NLPR) and the University of Karlsruhe have made important contributes to

47、 3-D model-based vehicle localization and tracking.</p><p>　　The research group at the University of Reading adopts 3-D wire-frame vehicle models. Tan et al. propose the ground-plane constraint (GPC), under

48、which vehicles are restricted to move on the ground plane. Thus the degrees of freedom of vehicle pose are reduced to three from six. This greatly decreases the computational cost of searching for the optimal pose. Moreo

49、ver, under the weak perspective assumption, the pose parameters are decomposed into two independent sets: translation parameters and r</p><p>　　The NLPR group has extended the work of the research group at t

50、he University of Reading. Yang et al. propose a new 3-D model-based vehicle localization algorithm, in which the edge points in the image are directly used as features, and the degree of matching between the edge points

51、and the projected model is measured by a pose evaluation function. Lou et al. present an algorithm for vehicle tracking based on an improved extended Kalman filter. In the algorithm, the turn of the steering wheel and<

52、;/p><p>　　The Karlsruhe group uses the 3-D wire-frame vehicle model. The image features used in the algorithm are edges. The initial values for the vehicle pose parameters are obtained from the correspondence

53、between the segments in an image and those in the projection model. The correspondence is calculated using viewpoint consistent constraints and some clustering rules. The maximum a posterior (MAP) estimate of the vehicle

54、 position is obtained using the Levenberg–Marquardt optimization technique.</p><p>　　The algorithm is data-driven and dependent on the accuracy of edge detection. Kollnig et al. also propose an algorithm bas

55、ed on image gradients, in which virtual gradients in an image are produced by spreading the Gaussian distribution around line segments. Under the assumption that the real gradient at each point in an image is the sum of

56、a virtual gradient and a Gaussian white noise, the pose parameters can be estimated using the extended Kalman filter (EKF). Furthermore, Haag et al. integrate</p><p>　　The above reviews model-based human bod

57、y tracking and model-based vehicle tracking. Compared with other tracking algorithms, model-based tracking algorithms have the following main advantages.</p><p>　　? By making use of the prior knowledge of th

58、e 3-D contours or surfaces of objects, the algorithms are intrinsically robust. The algorithms can obtain better results even under occlusion (including self-occlusion for humans) or interference between nearby image mot

59、ions.</p><p>　　? As far as model-based human body tracking is concerned, the structure of human body, the constraint of human motion, and other prior knowledge can be fused. </p><p>　　? As far a

60、s 3-D model-based tracking is concerned, after setting up the geometric correspondence between 2-D image coordinates and 3-D world coordinates by camera calibration, the algorithms naturally acquire the 3-D pose of objec

61、ts.</p><p>　　? The 3-D model-based tracking algorithms can be applied even when objects greatly change their orientations during the motion. Ineluctably, model-based tracking algorithms have some disadvantag

62、es such as the necessity of constructing the models, high computational cost, etc.</p><p><b>　　中文翻譯：</b></p><p>　　運動圖像和運動矢量檢測綜述</p><p>　　經(jīng)過運動檢測，監(jiān)控系統(tǒng)一般會在圖像序列中一幀一幀地跟蹤著運動目標

63、。跟蹤算法在處理過程中通常會與運動檢測有很大的交叉。跟蹤通常需要使用諸如點，線或塊特征來匹配連續(xù)幀的目標。常用的數(shù)學(xué)工具有卡爾曼濾波，Condensation算法，動態(tài)貝葉斯網(wǎng)絡(luò)，測地線法等。跟蹤方法可分為四大類：基于區(qū)域的跟蹤，基于動態(tài)輪廓的跟蹤，基于特征跟蹤和基于模型的跟蹤。應(yīng)當指出，這種分類不是絕對的，在不同類別的算法可以集成在一起。 </p><p>　　A、基于區(qū)域的跟蹤 </p><

64、;p>　　基于區(qū)域的跟蹤算法跟蹤目標是以移動目標對應(yīng)的圖像區(qū)域變化為根據(jù)的。對于這些算法，背景圖像是保持動態(tài)的，運動區(qū)域通常由背景相差法檢測到。Wren等。利用小區(qū)域特征進行室內(nèi)單人的跟蹤。文中將人體看作由頭、軀干、四肢等身體部分所對應(yīng)的小區(qū)域塊所組成， 利用高斯分布建立人體和場景的模型， 屬于人體的像素被規(guī)劃于不同的身體部分， 通過跟蹤各個小區(qū)域塊來完成整個人的跟蹤。最近，McKenna等在中利用色彩和梯度信息建立自適應(yīng)的背景模

65、型， 并且利用背景減除方法提取運動區(qū)域， 有效地消除了影子的影響； 然后， 跟蹤過程在區(qū)域、人、人群三個抽象級別上執(zhí)行， 區(qū)域可以合并和分離， 而人是由許多身體部分區(qū)域在滿足幾何約束的條件下組成的， 同時人群又是由單個的人組成的， 因此利用區(qū)域跟蹤器并結(jié)合人的表面顏色模型， 在遮擋情況下也能夠較好地完成多人的跟蹤。至于基于區(qū)域的車輛跟蹤方面也有一些典型的系統(tǒng)，像美國聯(lián)邦航空管理局（聯(lián)邦公路管理局）CMS動員系統(tǒng)支持系統(tǒng)和由Berkele

66、y組開發(fā)的噴氣推進系統(tǒng)等實驗室（JPL）和PATH系統(tǒng)。 </p><p>　　雖然他們在只包含幾個對象（如公路）工作很有成效，基于區(qū)域的跟蹤算法不能可靠地處理對象之間的遮擋的情景。此外，由于這些算法只能獲得在該區(qū)域的跟蹤結(jié)果，實質(zhì)上是對運動檢測，輪廓或三維程序構(gòu)成的對象不能被實現(xiàn)。（三維目標的構(gòu)成包括了定位與定向）。因此，這些算法不能滿足嚴密監(jiān)視零亂的背</p><p>　　景或多個移動

67、對象的要求。 </p><p>　　B. 基于活動輪廓的跟蹤</p><p>　　基于活動輪廓的跟蹤算法是利用封閉的曲線輪廓來表達運動目標， 并且該輪廓能夠自動連續(xù)地更新]。這些算法的目的是直接提取形狀的科目，并提供比基于區(qū)域算法更有效的描述。 Paragios等在中利用短線程的活動輪廓， 結(jié)合Level Set 理論在圖像序列中檢測和跟蹤多個運動目標。Peterfreund采用基于卡爾曼

68、濾波的活動輪廓來跟蹤非剛性的運動目標。Isard等在中利用隨機微分方程去描述復(fù)雜的運動模型， 并與可變形模板相結(jié)合應(yīng)用于人的跟蹤。Malik等在中已成功地應(yīng)用于主動輪廓的方法進行車輛跟蹤。相對于以區(qū)域為基礎(chǔ)的跟蹤算法，主動輪廓線的算法描述對象更加簡便，更有效地降低計算復(fù)雜度。即使在干擾或部分遮擋，這些算法仍可連續(xù)跟蹤對象。但是，跟蹤精度是有限的輪廓水平。在三維狀態(tài)下，從它在圖像平面輪廓對象的恢復(fù)是一個要求很高的問題。另一個困難是，活動輪

69、廓的算法是高度敏感的跟蹤初始化，因此初始化跟蹤很困難的。 </p><p>　　C.基于特征的跟蹤 </p><p>　　基于特征的跟蹤算法通過對象提取的元素進行識別和跟蹤，集群向更高層次的功能，然后匹配圖像間的特征?；谔卣鞯母櫵惴梢愿鶕?jù)所選功能性質(zhì)進一步分為三個類別：基于整體特征的算法，基于局部特征的算法，和基于依賴圖形的算法。 </p><p>　　?基于

70、整體特征的算法的特征包括質(zhì)心，周長，面積和顏色等。Polana等在[33]中提供一個很好的基于整體特征的跟蹤例子。將每個人用一個矩形框封閉起來， 封閉框的質(zhì)心被選擇作為跟蹤的特征； 在跟蹤過程中若兩人出現(xiàn)相互遮擋時， 只要質(zhì)心的速度能被區(qū)分開來，跟蹤仍能被成功地執(zhí)行。 </p><p>　　?基于局部特征的跟蹤算法的特征包括線段，曲線段和角頂點等。 </p><p>　　?基于圖形的算法的

71、特征包括各種間距和特征間的幾何關(guān)系。 </p><p>　　以上三種方法可以結(jié)合起來。Jang在中利用區(qū)域的形狀、紋理、色彩和邊緣特征信息建立了活動模板， 結(jié)合卡爾曼濾波的預(yù)測方法， 使特征匹配能量函數(shù)最小化來完成運動目標的跟蹤過程， 該活動模型對于非剛性目標的跟蹤具有很好的自適應(yīng)性。</p><p>　　一般來說，根據(jù)他們對二維圖像的作用，基于特征的跟蹤算法可以成功地運用并能迅速進行實時

72、處理跟蹤像高速公路需要多目標的場景等，但由于他們需要更多時間來搜索和圖形匹配，基于圖像跟蹤的算法不能用于實時跟蹤?；谔卣鞯母櫵惴梢蕴幚硗ㄟ^使用目標運動信息，局部特征和依賴圖表方法處理部分遮擋。然而，基于特征的跟蹤算法也有幾個嚴重缺陷。 </p><p>　　?隨著透視投影非線性失真和運動視點變化，基于2D目標特征的識別率會變得很低。 </p><p>　　?這些算法一般無法恢復(fù)三維目

73、標形狀。 </p><p>　　?在處理遮擋的穩(wěn)定性中，相互重疊和不相關(guān)性上性能很差。 </p><p>　　D.基于模型的跟蹤 </p><p>　　基于模型的跟蹤算法跟蹤目標是通過項目對象模型與圖像數(shù)據(jù)匹配事先預(yù)測。這些模型通常是由離線手工測量，CAD工具或計算機視覺技術(shù)路線構(gòu)建。作為基于模型的剛性目標跟蹤和基于模型非剛性目標跟蹤有很大不同，我們分別研究基于模型

74、的人體跟蹤（非剛性目標跟蹤）和基于模型的車輛追蹤（剛性目標的跟蹤）。 </p><p>　　1.基于模型的人體跟蹤：對于基于模型的人體跟蹤的一般方法被稱為分析的合成，它被用于預(yù)測匹配更新類型。首先，根據(jù)事先預(yù)測和全程軌跡預(yù)測下一個幀，然后，綜合預(yù)測模型和預(yù)測到的圖像數(shù)據(jù)與圖像平面進行比較。通過一個特殊的估值函數(shù)來衡量預(yù)測模型和圖像數(shù)據(jù)之間的相似性。根據(jù)不同的搜索策略，或是遞歸或是抽樣方法，直到找到正確的構(gòu)成并用于

75、更新這個模型。對第一幀估計需要特殊處理。一般來說，基于模型的人體跟蹤涉及三個主要問題： </p><p>　　?人體模型的構(gòu)成； </p><p>　　?運動模型和運動約束的事先預(yù)測的表示形式； </p><p>　　?預(yù)測和搜索策略。在這三個問題上的前人工作是簡要的，并分別回顧如下。 </p><p>　　A.人體模型：人體模型的構(gòu)建是基于

76、人體跟蹤。一般來說，越復(fù)雜的人體模型會得到越準確的跟蹤結(jié)果，但卻需要更高的計算成本。一般情況下，人體幾何結(jié)構(gòu)可以由以下四個方式來表示。 </p><p>　　?線圖法。人類運動的本質(zhì)通常包括軀干，頭部和四肢，因此線圖法是通過線和環(huán)節(jié)連接關(guān)節(jié)支來代表人體各部分。 Karaulova等在[25]中用線圖法使用隱馬爾可夫模型（HMM模型）建立了人體運動學(xué)的分層模型，實現(xiàn)了在單目圖像序列中跟蹤人體的獨立視點。</p

77、><p>　　?二維輪廓。該方法的使用直接與人體在圖像中的投影有關(guān)，如Ju等在[26]中提出了一種紙板人模型，它將人的肢體用一組連接的平面區(qū)域塊來表達， 該區(qū)域塊的參數(shù)化運動受關(guān)節(jié)運動(articulated movement) 的約束。 Niyogi等在[27]中利用在XYT空間使用空間時間模式來跟蹤，分析和識別移動的目標。利用時空切片方法進行人的跟蹤: 首先觀察由人的下肢軌跡所產(chǎn)生的時空交織模式， 然后在時空域中

78、定位頭的運動投影，接下來識別其它關(guān)節(jié)的軌跡，最后利用這些關(guān)節(jié)軌跡勾畫出一個行人的輪廓.</p><p>　　?立體模型。二維模型主要缺點是它們受視角限制。為了克服這一缺點，許多研究人員利用廣義錐臺、橢圓柱、球、二次曲面等三維模型來描述人體的結(jié)構(gòu)細節(jié)， 因此要求更多的計算參數(shù)和匹配過程中更大的計算量。Rohr使用14 個橢圓柱體模型來表達人體結(jié)構(gòu)。Wachter等在中利用橢圓錐臺建立三維人體模型。 </p&g

79、t;<p>　　?層次模型。 Plankers等在中為實現(xiàn)更精確的結(jié)果提出層次人體模型。它包括四個層次：骨架，橢圓球體模擬脂肪粒，多邊形表面代表皮膚和陰影渲染。 </p><p>　　B.運動模型：人類的四肢和關(guān)節(jié)運動模型常被用來跟蹤。由于肢體動作嚴格受到限制，它們還是很有效的。這些事先預(yù)測的運動模型是用來預(yù)測運動參數(shù)，去解釋和識別人的行為，或是限制低層次的圖像測量估計。例如，Bregler分解成多

80、個抽象的人類行為，并在低階段興建HMM模型代表高水平抽象用于跟蹤和識別。Zhao等人[106]提出用最小描述長度為芭蕾舞高度結(jié)構(gòu)化的運動模式功能（MDL）的范例。這個模型的模式類似于有限狀態(tài)機（FSM）。多變量主成分分析（MPCA方法）是用來訓(xùn)練Sidenbladh等人走路的模式。同樣，Ong等。采用分層PCA學(xué)習(xí)他們的運動模式，是基于在全球eigensapce不同子空間上的轉(zhuǎn)移概率矩陣和全局特征空間之間的轉(zhuǎn)移概率矩陣。Ning等在文獻

81、[7]中，汲取半自動獲得訓(xùn)練樣例和使用高斯分布來表示模型。 </p><p>　　搜索策略，對那些在高維體配置空間估計上是很困難的，所以，搜索策略往往需精心設(shè)計，以減少解空間。一般來說，有主四個要類別的檢索策略：動態(tài)，泰勒模型，卡爾曼濾波和隨機抽樣。動態(tài)策略的使用適用于被跟蹤對象的剛性目標的三維模型。作為探視信息，它主要是平衡三維模型和造成的實際對象之間構(gòu)成差異最小化。對泰勒模式為基礎(chǔ)的戰(zhàn)略，逐步改善現(xiàn)有的估計，

82、使用觀察運動參數(shù)的差異來預(yù)測更好的搜索方向。雖然它能發(fā)現(xiàn)局部極小值，但不能保證找到全局最低。作為一個遞歸線性估計，卡爾曼濾波能徹底處理了在相對混亂的實時跟蹤的形狀和位置的運動參數(shù)的密度，并可以良好的好以高斯建模。為了處理運動參數(shù)的概率密度函數(shù)的多式聯(lián)運和非高斯引起混亂，隨機取樣策略，如馬爾可夫鏈蒙特卡洛，遺傳算法，和壓縮算法，都設(shè)計各種假說解決這些問題。其中隨機取樣的視覺跟蹤戰(zhàn)略和壓縮是最流行的。 </p><p&g

83、t;　　基于模型的車輛追蹤：對于基于模型的車輛跟蹤，主要是使用三維線框車輛模型[95]。雷丁大學(xué)（University of Reading）研究小組，在模式識別國家重點實驗室(NLPR)和德國卡爾斯魯厄大學(xué)（University of Karlsruhe）大學(xué)的研究小組對基于三維模型為基礎(chǔ)的車輛定位和跟蹤方面做出了重要貢獻。 </p><p>　　雷丁大學(xué)研究小組采用了三維線框車型。在文獻中，Tan等根據(jù)該車輛

84、只限于在地面上移動，提出了地平面約束（GPC）。因此，車輛的自由度從6個降低到3個。這大大降低了為尋找最優(yōu)姿勢的計算成本。此外，在弱透視投影假設(shè)下，構(gòu)成參數(shù)分解為兩個獨立的部分：平移參數(shù)和旋轉(zhuǎn)參數(shù)。Tan等在中提出了一種廣義霍夫變換算法來估計車輛的構(gòu)成部分的一個特征線段。此外，Tan等在[121]中分析了一維圖像梯度關(guān)系，并通過表決確定該車輛形狀。至于車輛形狀的改進，雷丁大學(xué)的研究小組已經(jīng)從過去的工作方法中找到一種利用獨立的一維搜索。最

85、近，Pece等在介紹估計車輛構(gòu)成統(tǒng)計牛頓法。 </p><p>　　在模式識別國家重點小組延長雷丁大學(xué)研究小組的工作。Yang等人在中提出了一種新的三維模型為基礎(chǔ)的車輛定位算法，直接將圖像中的邊緣點作為特征，邊緣點和預(yù)測模型匹配程度之間的自由度是通過評價函數(shù)測量的。Lou等在[174]中提出一種基于擴展卡爾曼濾波車輛跟蹤算法的改進算法。在算法中需要考慮方向盤的轉(zhuǎn)向與前后輪距離。由于這跟控制車輛的運動的駕駛員和假定

86、動態(tài)模型有直接關(guān)系，在進行復(fù)雜的動作行為時改進的擴展卡爾曼濾波性能優(yōu)于傳統(tǒng)的擴展卡爾曼濾波的性能。 </p><p>　　Karlsruhe 小組采用三維線框車型。在圖像的算法使用邊緣作為特征。車輛初始值是獲得在圖像和預(yù)測模型中各部分之間的通信。通信是依靠使用視點一致約束和一些聚類規(guī)則。車輛最高后驗（MAP）的位置，估計是包含使用的Levenberg - Marquardt優(yōu)化技術(shù)。 </p>&l

87、t;p>　　該算法是數(shù)據(jù)驅(qū)動的和取決于邊緣檢測的準確性。Kollnig等在文獻中也提出了基于圖像梯度的一種算法，它是一種在圖像中產(chǎn)生的虛擬梯度散布周圍的線段高斯分布的算法。根據(jù)假設(shè)，即在每一個像點真正梯度是一個虛擬的梯度和高斯白噪聲之和，可估計造成使用擴展卡爾曼濾波器（EKF）的參數(shù)。此外，Haag等在中綜合Kollning等基于圖像梯度與基于光流的算法，該方法使用的圖像功能的社區(qū)評估圖像梯度。然而，光流上使用圖像特征的全球信息，

88、橫跨利息率（ROI），整個地區(qū)的一體化。因此，梯度和光流是信息的補充來源。 </p><p>　　以上回顧了基于模型的人體跟蹤和基于模型的車輛跟蹤。相比其他跟蹤算法，基于模型的跟蹤算法具有以下主要優(yōu)點。 </p><p>　　?通過使對三維輪廓或?qū)ο蟮谋砻孢M行事先預(yù)測，該算法在本質(zhì)上是強健的。即使在（包括自身遮擋）或附近的圖像運動之間的干擾的情景下，該算法可以取得更好的效果。 </p

89、><p>　　?就基于模型的人體跟蹤而言，人體結(jié)構(gòu)，人體運動的限制，和其他事先預(yù)測可以融合。 </p><p>　　?就基于三維模型的追蹤而言，通過攝像機標定坐標建立對應(yīng)關(guān)系二維幾何圖像坐標和三維全局坐標，算法自然需要目標的三維形態(tài)。 </p><p>　　?即使在目標有極大地方向改變的運動情形下，基于三維模型的跟蹤算法也可應(yīng)用。當然，基于模型的跟蹤算法也有一些缺點如在

90、建模必要性和高計算成本等。</p><p>　　作者：Weiming Hu, Tieniu Tan, Fellow, IEEE, Liang Wang, and Steve Maybank</p><p>　　出處：IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL