A. Blake and M. Isard, Active Contours. Springer-Verlag, 1998.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....Deformable Model Deformable contours[9] are contour and surface extraction tracking tools based on the minimization of a contour energy that defines the desired contour properties and interesting image features. In order to minimize the contour energy, the originators and many other researchers[4, 5, 7] used a gradient descent algorithm, which has problems with local minima and sensitivity to initial contour positions. In order to address these problems, an energy minimization method based on dynamic programming was proposed[3, 6] that guarantees a global minimum within the defined search ....

....V vanish. Therefore, the gradient direction for the minimization search a#ects only the j positions of each m if , which is represented by jOf(m if ) Consequently, the search vector for the negative of the gradient direction is obtained by D = #EMesh (M) #jOf(m if ) T (4) where f =1. F and i =1. N.NotethatD is a NF dimensional vector and there are 2FN number of mesh elements. Intuitively, D assigns a one dimensional direction vector for each mesh element. The vectors can be towards the positive or negative j directions, or they can be 0. We execute the following ....

[Article contains additional citation context not shown here]

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

....a general (infinite dimensional) deformation. Our aim is to define tracking as a trajectory on a finite dimensional group, despite infinite dimensional deformations. Substantially di#erent in methods, but related in the intent, is the work on stochastic filters for contour tracking and snakes (see [8] and references therein) There, however, what is being tracked over time is a general deformation (although finitely parametrized via splines or other parametric descriptions) rather than a (group) motion. Therefore, the end product of these tracking algorithms is not a trajectory on a ....

A. Blake and M. Isard. Active contours. Springer Verlag, 1998.

....of Rochester, and demonstrated to work robustly with a small collection of ordinary 3D objects with multiple cameras surveying multiple rooms. 2. Previous Work Tracking involves following a set of features through a sequence of images. Various different features have been used, such as contours [3], corners [20] shape and color [25] Features detected in one image are matched to corresponding features in the previous image. Feature detection from one image to the next is frequently unreliable due to noise and occlusion. To handle this problem, researchers have employed techniques such as ....

....to corresponding features in the previous image. Feature detection from one image to the next is frequently unreliable due to noise and occlusion. To handle this problem, researchers have employed techniques such as the Kalman lter [1] and conditional density propagation (Condensation) algorithm [3]. Both these algorithms work by estimating, or predicting, the location of the features being tracked in the next image, and using the error between the predicted and measured location to update the predictive model. Image based object recognition techniques are more general and more easily ....

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

.... requires precise camera calibration and the computation of non trivial features [28] In this paper, we define a mixedstate model in which (i) human heads in the image plane are modeled as elements of a template space, allowing for the description of a template and a set of valid transformations [2], and (ii) cameras depicting people are indexed by a discrete variable. Specifically, a state is defined by X t = k t , x t ) k 1 , x t where k t is a discrete NK valued camera index, and x t is a continuous vector in the space of transformations . Furthermore, the dynamical model ....

...., and the weighted mean of the continuous component x t given the MAP discrete estimate are computed by k t = arg max i#I j t ; x t = t x , 3) where j = i k t = j . 2. 2 Person model Speaker heads are represented by their silhouettes (contours) in the image plane [2]. In particular, we used a parameterized vertical ellipse to represent the basic shape. 2.3 Dynamical models The uncountable set of TPMs Tmn (x t 1 ) is coarsely quantized based on the value of (T t 1 , T t 1 ) to allow for camera switching based on the speaker location at the previous ....

[Article contains additional citation context not shown here]

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

....interface applications such as sign language recognition and visually mediated interaction. Existing methods for markerless tracking can be categorised according to the measurements and models used [9] In terms of measurements, tracking usually relies on intensity information such as edges [10, 2, 17, 5], skin colour and or motion segmentation [24, 14, 11, 16] or a combination of these with other cues including depth [13, 25, 19, 1] The choice of model depends on the application of the tracker. If the tracker output is to be used for some recognition process then a 2D model of the body will ....

Andrew Blake and Michael Isard. Active Contours. Springer-Verlag, 1998.

....74] are needed for more complicated scenes or longer ones. Note that due to the similarity between frames, our algorithm converges very fast (typically in less than 30 iterations, insuming a few seconds on our PC, of a non optimal explicit implementation) The CONDENSATION algorithm described in [12] can also achieve, in theory, topology free tracking, though to the best of our knowledge real examples showing this capability have not been yet reported. In addition, this algorithm requires having a model of the object to track and a model of the possible deformations, even for simple and ....

.... is very general and allows for improvement, specially in the Morphing Active Contours algorithm: finding more robust selections of the feature map and the discrepancy function , using more than just 2 frames (via Kalman filtering or using the techniques in the novel scheme developed in [12]) proving existence and uniqueness theoretical results. 5.2 Image Inpainting In Chapter 3 we introduced a novel algorithm for image inpainting that attempts to replicate the basic techniques used by professional restorators. The basic idea is to smoothly propagate information from the ....

A. Blake and M. Isard, Active Contours, Springer-Verlag, New York, 1998.

....ally mediated interaction, correctly interpreting and tracking multiple human body parts are critical. Reliable markerless tracking of the human head and hands configuration is often a pre condition for natural gesture recognition. Such tracking usually relies on imagery information such as edges [18,3,38,8], skin colour and motion segmentation [57,32,24,35] or a combination of these with other cues including depth [29,59,41,1] If tracking is to be used for recognition, a 2D model of the body will suffice [35,24] On the other hand, a 3D model of the body may be required for generative purposes, ....

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

.... which violates the independence assumption implicit in the measurement model (Gaussian sampling noise with variance r) The formal solution to this problem is to use the covariance matrix R = rH , where H ij = Z H i (u)H j (u)du and H i is the interpolation function for the parameter (see [60]) In practice a covariance matrix S is estimated from training data to obtain 2n eigenshapes using the relationship SHe i = i e i . The eigenvectors e i are orthogonal to the inner product hu; ui = u Hu. Orthonormality is enforced by scaling these vectors appropriately. The shape vectors ....

....from the model in order to generate a set of more or less likely continuations. The model can also ll in missing parts of the behaviour interaction using a Bayesian framework in a similar manner to representing and updating the a posteriori density used in the tracking algorithm [60]. In general, the set of plausible state hypotheses can be found from this density function, where the maximum represents the most likely hypothesis for the state of the interaction. The main application of their work is in the synthesis of a virtual partner and this has been demonstrated for ....

A. Blake and M. Isard, Active Contours, Springer-Verlag, 1998.

....Then, semantic object segmentation can be achieved by change detection and or motion segmentation [1] Other methods impose constraints on the shape of the tracked object by using 2D [2] 3] or 3D shape [4] 5] models. Some algorithms acquire the 2D shape space information through training [6], 7] and use projections onto the shape space to estimate the most likely object boundary at a certain frame. The CONDENSATION algorithm [8] which is a state space sampling approach, needs the shape space to be known beforehand. Another ap plication of learned motion models for tracking is ....

A. Blake and M. Isard, Active Contours, Springer-Verlag, 1998.

....ROBABILISTIC visual contour tracking has been an active research area in the computer vision community in the last ten or more years. It has many potential applications in intelligent robots, in monitoring and surveillance, in biomedical image analysis, in human computer interfaces, etc. [1]. For these tracking tasks, a common approach is the use of the Kalman filter or extended Kalman filter. While some researchers employ physical snakes as system models in the (extended) Kalman filter [2,3,4] others use constant velocity motion models or learned motion models from training image ....

.... between curves, L 2 norm for the curve r(s) is used, and accordingly, norms Ilollfor control point vector Q and shape vector can be defined as IIQII = IIA = IIr s 11 where L 1 IIr(x)11= r(s)rr(s)ds) 2 , IIQII= QUQ s = 0 1 Details about the definitions of the norms can be found in [1]. In many situations, the shape space is chosen as plan affine space where shape space can be described as 0 0 i 0 The first two columns of w represent horizontal and vertical translation, the last four columns represent rotation, scaling and shearing. In practice, we choose Q0 to have its ....

A. Blake, M. Isard, Active Contours. Springer-Verlag, 1998.

....and (2) have no analytical solution. A common solution is thus the approximation of the posterior p(# t 1:t ) and to solve this numerical problem with a computationally e#cient algorithm. Sequential Monte Carlo methods [12] also known as particle filtering, bootstrap filtering or Condensation [15] have proofed to be suitable for this purpose. 2.1 Knowledge Hierarchy The complexity of the problem increases non linearly with every additional object, since object compounds are not only able to model the state of individual objects but also their interactions and group behavior. Bayesian ....

Blake, A., Isard, M.: Active Contours. Springer Verlag (1998)

....a general (infinite dimensional) deformation. Our aim is to define tracking as a trajectory on a finite dimensional group, despite infinite dimensional deformations. Substantially di#erent in methods, but related in the intent, is the work on stochastic filters for contour tracking and snakes (see [6] and references therein) Our framework is designed for objects that undergo a distinct overall global motion while locally deforming. Under these assumptions, our contribution consists of a novel definition of motion for a deforming object and a corresponding definition of shape average ....

A. Blake and M. Isard. Active contours. Springer Verlag, 1998.

....3D pose model with optical flow information from different viewpoints. The authors of [17] stress that adaptivity of the applied observation models is of crucial importance if tracking ought to be stable even in cluttered environments. ICONDENSATION [6] a derivative of the popular CONDEN SATION [1] algorithm exploits importance sampling for improving the performance and stability of particle filter based tracking. The algorithms proposed in [7] 5] 11] and [15] have in common that they all solve the multi object tracking problem by substituting the single object models with compounds ....

A. Blake and M. Isard. Active Contours. Springer Verlag, 1998.

....4.1 Spline contours, configuration space Let us consider the spline curve s : 0; L] R defined by its support points q = q ; q = i. e: s(t) A (t) A (t) where (t) 1 (t) 2 (t) n (t) are the usual Bspline basis functions (cf. [2]) and s(i) q . A is linear transformation mapping control points to support points and depends only on . Let q 0 = q ; q ) be a vector of support points defining the contour s 0 . If G is a group of similarity transformations of the 2D plane (R ) then one can find a ....

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

....as shown in Figure 2(b) The mask created was clearly inadequate for our application, and although operated in real time was simply ineffective. a) b) Figure 2. a) Composite frame bimodal histogram, b) binary threshold mask superimposed on frame 1. Active contour models (ACMs) or snakes [10] have been implemented by Kass et al. 11] Cohen [12] Xu and Prince [13] but few (Cheng et al. 14] Levienaise Obadia and Gee [15] have applied ACMs to ultrasound images for object location and tracking. Most of these authors have noted problems tracking features in high noise images due to ....

....the tendon to assess the ACM s performance. The final ACM result shown in Figure 3(b) demonstrates poor convergence in several places including the lower skin layer and upper finger flexor, making this approach somewhat impractical for our application. We intend to investigate CONDENSATION [10] as a future means of tracking the tendon. a) b) Figure 3. a) Initial snake input points, b) final converged snake output. Having exhausted many techniques found in the literature and due to the straight like characteristics of the tendon, we implemented the standard HT [21] to detect ....

[Article contains additional citation context not shown here]

A. Blake & M. Issard. Active Contours. Springer-Verlag, New York, 1998.

....(red standing for low and green for high) b) 3D vector mask after thresholding. 4.5 Building 3D strands We chose to build the final 3D hair strands in two steps. First, we build chains of pixels in image space, using the 2D vector field of the sequence mask and using the snakes technique [12, 2]. Each node of each pixel chain is related to a 2D vector, itself associated to a 3D vector (see section 3.3) We finally get the third dimension, in the building process, by using the information given by 3D vectors of each node. 5 Results Our system allows the extraction of hair strand ....

A. Blake and M. Isard. Active Contours. Springer-Verlag 1998.

....of low dimensional linear or multi linear models from training data has become a standard paradigm in computer vision. A variety of linear learning models and techniques such as Principal Component Analysis (PCA) 20, 37, 38, 44, 67] Factor Analysis (FA) 22, 44] Autoregressive analysis (AR) [9], and Singular Value Decomposition (SVD) 28] have been widely used for the representation of high dimensional data such as appearance, shape, motion, temporal dynamics, etc. These approaches differ in their noise assumptions, the use of prior information, and the underlying statistical models, ....

A. Blake and M. Isard. Active Contours. Springer Verlag, 1998.

.... (e.g. in the sequences that we tried, the face can move more than 20 pixels from frame to frame) In order to cope with such real conditions, we explore the use of stochastic methods such as Simulated Annealing (SA) 2] Genetic Algorithms (GA) 27,29] or Condensation (particle filtering) [6,17] for motion estimation. Although the techniques are very similar computationally speaking, here we make use of GA [29] within a coarse to fine strategy. Given the first image of the sequence, we manually initialize the layers or masks at the highest resolution level and assign the graylevel to ....

A. Blake and M. Isard. Active Contours. Springer Verlag, 1998.

....process at and the velocity of the pedestrian, but is undoubtedly smaller than the whole image itself. Also, the more accurate this prediction step is, the smaller the region where the pedestrian could be and hence the less shape parameters needed to detect the pedestrian. The active contour model [24, 16, 25] has shown to triumph in the area of non rigid object tracking due to its deformable nature and has solved the tracking problem eciently for many domains. 19 The initial phase of the pedestrian detection process, i.e. pre tracking, assumes no prior knowledge of the position of the pedestrian in ....

A. Blake and M. Isard, Active Contours. Springer-Verlag, 1998.

....can be used to describe the location and velocity of other limbs. State vectors containing information on different limbs could be stacked to form a single state vector and the # and # matrices stacked diagonally to form the new coefficients to the kinematic equation. Many current systems [3, 9] do not attempt to model kinematic motion and instead rely on the process noise, # , to account for the change in limb position from one frame to the next. British Machine Vision Conference 4 Other motion capture systems use complex models in which further parameters are required to define the ....

....and instead rely on the process noise, # , to account for the change in limb position from one frame to the next. British Machine Vision Conference 4 Other motion capture systems use complex models in which further parameters are required to define the state s expected trajectory through time [11, 3]. These extra parameters are contained in the matrices # and # and often need to be learnt from a series of training data. Such models are therefore restrictive because they limit the types of motion which can be tracked. To overcome this limitation, these systems are usually forced to use ....

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

....the segmentation results which contain the information at only one time step are not sufficient to establish a safe control system. To increase the robustness, a tracking process which can handle multiple hypotheses has to be included. Our work has evolved from the CONDENSATION algorithm [3, 8, 9] developed for contour tracking in visual clutter. Outlines and features of moving foreground objects, modeled as curves, are tracked in video sequences. Some elements in the background clutter may consist of objects similar to the foreground object, for instance when a person is moving past a ....

....this paper we first introduce the mathematics needed to formulate the CONDENSATION algorithm. Secondly, we explain how it can be extended to track multiple objects and to cope with newly appearing objects and present applications. Finally, a discussion shows the differences to the original scheme [3, 8, 9] and also compares our approach to Kalman filtering [4, 7] 2. Mathematical Methods The notation in this paper approximately follows [3, 8, 9] The terminology is listed in Fig. 1. Probability Distribution: An object is characterized by a state vector x 2 X. Assuming that we are not able to ....

[Article contains additional citation context not shown here]

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

....are from each other; and the level of occlusion of the target object. Once a snake loses focus from the target object, it rarely regains the successful detection and tracking of that target object. A more detailed discussion of active contour models and their energy functions can be found in [8] and [9] 2 . 2 Tracking Walking Humans In order to produce a clean and reasonably varied set of data, the 3D modelling and animation package Poser [10] was used to simulate human movement. Snakes were relaxed around simulated humans in 30 different movies, all of which contained a single ....

....the results sections of this paper. The target human has been dimmed in this figure to increase the snake s visibility. Figure 2 shows the results of a snake tracking a sample human over 120 frames of video footage. No objective measure exists of an active contour s success at tracking objects [8], but here the model was clearly successful as the humans were tracked in all of the 30 different instances. Speed is not a primary concern in this work, but it is worth noting that the relaxation times for the snakes in moving from one frame to another was often very fast. As can be seen from ....

. Blake A. & Isard M. [1998]. Active Contours. Springer-Verlag.

....has been detected, the information about its position, orientation and size is passed to the tracker module. As shown in fig.5, this module comprises an estimator, a controller and a measure module connected in the conventional closed loop fashion commonly adopted for visual object tracking [1]. At each frame the Kalman Tracker, on the basis of the previous observations (measures) produces an estimate of the new status of the fingertips, the accuracy of which tends to improve at each iteration (in the ideal case, the error tends to zero) thanks to the information provided by each new ....

....closely related to the hypothesis that both the noise vectors and the status vector have a Gaussian distribution. At this stage we will not address this issue, since the performance of the Kalman tracker is reasonable for our purposes; several di#erent solutions do, however, exist for this problem [1]. The measure module basically uses the same detection algorithm as is used for the initialisation phase, described in Fig.4. In this case, however, it is applied to a rectangular area B for each fingertip, centred in the position predicted by the tracker as shown in Fig.6 and of a size ....

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

....the matching between hypotheses and image observations. For example, template matching tracking method often takes SSD as the measurement. The evaluation would be quite challenging when measuring a shape hypothesis in a clutter background. Although some analytical results were reported in [3], many current tracking methods take ad hoc measurements. Hypotheses generating is to produce new hypotheses based on old estimation of target s representation and old observation. Target s dynamics could be embedded in such a predicting process. Intuitively, hypotheses generating characterizes ....

....and an a#ne transformation. The shape samples in our algorithms are drawn in the shape space, i.e. X s = A 11 , A 12 , A 21 , A 22 , t 1 , t 2 ) 5.2 Shape Observation It is crucial to have an accurate shape observation in tracking. Our implementation takes a similar approach used in [3]. Edge detection is performed in 1 D along the normal lines of the hypothesized shapes, shown in Figure 5. Thus, observation reduces to a set of scalar positions z = z 1 , z M ) due to the presence of clutter. The true observation z could be any one of them. So, p(z x) qp(z clutter) ....

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

....snake is an image feature technique that uses energy terms defined from gradient features of global interest to find the desired contour [7] In most real applications such assumption is too strong. Different authors suggest to combine the gradient based potential with valley and crest maps [16, 3]. However, on one hand, the best way of integrating different features remains to be an open This work is supported by CICYT and EU grants TIC98 1100 and 2FD97 0220, and Xunta de Galicia grant PGIDT99PXI20606B. problem and in the other hand, these features are not selective enough yet. This leads ....

.... modeling features of faces [16] More general methods, as Fourier descriptors, have been used for representing shapes in medical images [14] Alternative approaches based on modal analysis have also been proposed to constraint the model to deform only in ways implied by the training set of shapes [4, 3]. Although shape models convey important information, they are not the panacea; high accuracy techniques must make the most of grey level information too. In line with this idea, shape models and appearance models are combined in face recognition [8] Our approach based on the statistics ot the ....

A. Blake and Isard. Active contours. Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard, Active Contours. Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

No context found.

Andrew Blake and Michael Isard. Active Contours. Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard, Active Contours, Springer-Verlag, New York, 1998.

No context found.

A. Blake and M. Isard. Active Contours. Springer Verlag, April 1998.

No context found.

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard, Active Contours, Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard, Active Contours, Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard, Active Contours, Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard, Active Contours. Springer Verlag, 1999.

No context found.

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard, Active Contours, Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard. Active Contours. Springer Verlag, 1998.

No context found.

A. Blake and M. Isard. Active Contours. Springer-Verlag, London, 1998.

No context found.

A Blake, M Isard, Active Contours, Springer-Verlag, Berlin 1998.

No context found.

A. Blake, Active Contours. Springer Verlag, 1998.

No context found.

A. Blake, and M. Isard. Active Contours. Springer-Verlag, London, 1998. 283

No context found.

Andrew Blake and Michael Isard. Active Contours. Springer-Verlag, London, 1998.

No context found.

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

No context found.

A Blake, M Isard, Active Contours, Springer-Verlag, Berlin 1998.

No context found.

A. Blake and M. Isard. Active Contours. Springer-Verlag, 1998.

No context found.

A. Blake and M. Isard, Active Contours, Springer-Verlag, 1998.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC