9512.net

甜梦文库

甜梦文库

当前位置：首页 >> >> # A Robust Multiple Hypothesis Approach to Monocular Human Motion Tracking

A Robust Multiple Hypothesis Approach to Monocular Human Motion Tracking

Cristian Sminchisescu INPG-INRIA Cristian.Sminchsiescu@inrialpes.fr Bill Triggs CNRS-INRIA Bill.Triggs@inrialpes.fr

model to image data association. In this context, the critical issues include strongly non-linear problem structure, observability and singularity issues, unconstrained motions, and the ambiguous assignment of model predictions to image observations, particularly in cluttered scenes and with inherently imperfect body and clothing modeling. The main contribution of this paper is a robust, computationally ef?cient algorithm for estimating the posterior distribution over the model parameters and propagating it through time. As in other sampling-based approaches, the distribution is represented by a set of hypotheses together with a hypothesis generation and focusing mechanism. Our novel focusing mechanism captures the uncertainty inherent in the generalized model-image mapping. Our hybrid cost optimization algorithm involves both continuous and discrete components, namely: (i) local continuous optimization based on robust error distributions and incorporating joint limits; (ii) informed discrete sampling and hypothesis evaluation concentrated on the high-uncertainty parameter combinations of the continuous estimate; and (iii) a ?nal hypothesis pruning and propagation stage. We also address the problem of semi-automatic model initialization from a single image, present a hierarchical algorithm for estimating body pose and proportions, and show experimental tracking results on video sequences involving complex motions with signi?cant background clutter and self-occlusion.

Abstract

We study the problem of articulated 3D human motion tracking in monocular video sequences. Addressing problems related to unconstrained scene structure, uncertainty, and the high-dimensional parameter spaces required for human modeling, we present a novel, layered-robust, multiple hypothesis algorithm for estimating the distribution of the model parameters and propagating it over time. We use cost function based on robust contour and image intensity descriptors in a multiple assignment data association scheme. Our mixed discrete/global and continuous/local search technique uses both informed sampling and continuous optimization. Its novel hypothesis generation and pruning strategy focuses attention on poorly constrained directions in which large parameter space deviations are most likely, thus adaptively tracking the complex cost surface produced by non-linear kinematics, perspective projection and data-association problems. We also address the issue of semi-automatic acquisition of initial model pose and proportions, and show experimental tracking results involving complex motions with signi?cant background clutter and self-occlusion.

1 Introduction

Extracting 3D human motion in realistic unconstrained environments based on monocular video sequences is a dif?cult process owing to the large number of parameters that even minimal generative human models necessarily have, the incomplete observability of monocular projection, and the ambiguities inherent in complex scene structure. We advocate an optimization based approach to this problem, the essence of which is hybrid local/global minimization of a cost function representing the posterior probability of the model given the data. Our cost function embodies the generalized model-image mapping that results from chaining the sequence of complex transformations that predict and evaluate model con?gurations in the image. It includes the effects of non-linear kinematics, perspective projection and 1

1.1 Relation to Previous Work

Approaches to body and body part tracking using 3D models can be broadly classi?ed as either deterministic continuous [7, 8, 18, 28, 27, 15], discrete [13, 19], or stochastic based on particle ?ltering [11, 10, 24] (see [14] for a comprehensive review). Continuous approaches typically use locally linearized model approximations and unimodal Gaussian error distributions, often propagated in time using extended Kalman ?ltering. Stochastic approaches use sample-based representations for the underlying distributions, priors based on various expected motion models, and search focusing methods ranging from importance sampling to annealing in an attempt to recover the ‘typical set’ of

probable con?gurations. As far as we know, no continuous approach has used either continuous optimization based on robust likelihood models, or optimizers incorporating hard joint limits (although [27] enforces joint limits at the level of the Kalman ?lter). Nor has any dealt with robustly extracted image descriptors, nor used a multiple assignment scheme, nor modeled or searched in a robust way the expected multi-modal distribution over parameters. Only a few works have addressed monocular, 3D tracking in a robust setting, particularly for full human body models. Deutscher [10] uses a MCMC technique based on annealing to sample the high dimensional space, but uses a clean background, an undressed person and silhouette and edge-based components to regularize his cost function. He tracks a walking person using annealing and temporal models but requires 3 cameras to track less periodic motion. Sidenbladh [24] uses a similar particle-based technique to track a walking person in a more complex setting, using an importance sampling method based on a strong learned prior motion model and an intensity-based cost function. Both of these approaches make strong assumptions, either on the simplicity of feature assignment (near-perfect silhouette and edge localization) and/or on the type of scene motions, to derive specialized imporance sampling functions. Although the use of temporal models is appealing, any tracker designed to handle an unconstrained environment necessarily has to deal with irregular behavior, at best requiring several motion models and managing transitions between them. Unfortunately, multiple-model estimation schemes in high-dimensional spaces are at present very expensive computationally [5]. While not underestimating the importance of incorporating prior knowledge of the problem domain, in the present work we focus on techniques rooted in the generative model approach. We isolate two essential components of any optimization scheme based on a generative model, be it deterministic or stochastic. The ?rst is the careful design of the cost function, which must robustly integrate the extracted image cues (contours, image intensity, motion boundaries) to limit the number of spurious local minima in parameter space. The second is the analysis of the generalized mapping linking model parameters to observations, which we believe is indispensable to allow accurate focusing of the search effort on the parameter space regions where good cost minima are most likely to occur. We argue that analysis of the uncertainty structure of this mapping is an effective generic technique for computationally tractable search in high-dimensional problems. The paper is organized as follows: x2 brie?y reviews the parameterization of our generative model; x3 presents the 2

design of the cost function in terms of robust error distributions and image descriptors; x4 describes our hybrid continuous/discrete optimization method; x5 presents our semiautomatic initialization method; x6 discuses experimental tracking results; and x7 summarizes and discusses future research directions.

2 Generative Model

Our articulated model consists of kinematic and volumetric components. The underlying skeleton of the body is modeled in terms of kinematic chains associated body parts, represented using minimal parameterizations ( xa ). The volumetric structure of the parts is modeled using superquadric ellipsoids, with additional tapering and bending parameters (xd ) for approximate surface modeling [1]. Each ellipsoid is discretized as a mesh in its topological domain (see Figure 1). A typical model has around 30 kinematic d.o.f. Supplementary parameters include internal proportions ( xi ) (8 parameters encoding the positions of hip, clavicle or skull tip joints and 9 parameters for each volumetric part). Together, these produce a rather large parameter space, but only the kinematic parameters are actually estimated during tracking, although some of the internal proportions and volumetric parameters are estimated during model initialization (see Section 5). We believe that such a model offers several advantages in terms of high-level interpretation and occlusion prediction, and a good trade-off between complexity and coverage. We encode the model representation in a single parameter vector x xa ; xd ; xi . Any node ui 2 corresponding to one of the model parts, can be transformed into a 3D point pi pi x , and subsequently into an image prediction ri ri x by means of a composite non-linear transformation:

=( = () = ()

)

ri = T (x) = P (pi = A(xa ; xi ; D(xd ; ui )))

(1)

where D represents a sequence of parametric deformations which construct the corresponding part in a self-centered reference frame, A represents a chain of rigid transformations on the corresponding part kinematic chain, and P represents the perspective projection of the camera. The process of model estimation involves a data association problem between individual model feature predictions ri and one or more observations, that we shall generically denote ri (with additional subscripts if these are several). We refer to ri x ri ? ri x as the feature prediction error.

( )=

()

total probability according to Bayes rule:

1 1 p(xjr ) = Z p(rjx)p(x) = Z exp(? p(ri jx)di)p(x) ( ) ()

Z

(4)

where pi ri jx is the probability associated with the observation i, p x is a prior on model parameters and Z a normalization constant. The probability for the observation i, expressed in terms of corresponding model prediction is pi x N1 pui x , where N is the total number of model nodes, Wi is a positive de?nite weighting matrix associated to the assignment i, and:

( )=

()

8 < pui (x) = :

1 2

i ( ri (x)Wi bf = ; occ = k ;

ri (x)> );

k 1

if i is assigned if back-facing < , if occluded (5)

Figure 1: Human body model used during tracking

3 Cost Function Design

Whether continuous or discrete, any optimization process is dependent on the cost function to be minimized. Besides smoothness properties, we believe that cost functions should be designed to limit the number of spurious local minima in parameter space. Our approach employs a combination of edge and intensity information on top of a multiple assignment strategy based on a weighting scheme that focuses attention towards motion discontinuities. We also aim for a probabilistically interpretable model and build our cost function around robust error distributions

f (x) = ? log(Prjx)p(x)) p( 1 = N1 ( 2 a a( ra (x)Wa ra (x)>) (6) +Nbf bf + Nocc occ) + fp(x) where fp (x) is the negative log-likelihood of the prior, while Nocc and Nbf the occluded and back-faced (selfoccluded) number of model nodes, respectively.

In our MAP approach, we discretize the continuous problem and attempt to minimize the negative loglikelihood for the total posterior probability, expressed as the following cost function:

3.2 Image Descriptors

We choose both edge and intensity features for our cost function design. The images are smoothed with a Gaussian kernel, then a Sobel edge detector is applied. We employ a robust multiscale optical ?ow computation [3] to obtain both a ?ow ?eld and an associated outlier map. The outlier map conveys useful information about the motion boundaries and is used to weight the signi?cance of edges (see Figure 2). The motion boundaries are processed similarly to obtain a smooth image. For each node on model occluding contour, we perform a line search along the normal and retain all possible assignments within the search window, weighting them based on both the edge strength and their importance quali?ed by the motion boundary map. For nodes lying inside the object, we use intensity information derived from the robust optical ?ow. Including multiple edge assignments and ?ow, the ?rst term in the cost function (6) becomes:

3.1 Error Distribution and Functional Form

Robust parameter estimates are intrinsically related to the choice of a realistic likelihood model that embodies the expected total inlier plus outlier distribution for the observation. We model the distribution in terms of robust radial terms, i , where i s can be any increasing function with d and ds i 2 , that models an error disi tribution corresponding to a central peak, with in?uence , and a widely spread background of outliers . In this work, we use the following two robust error distributions (usually known as ‘Lorentzian’ and ‘Leclerc’ [4]):

(0) = 0

() (0) =

i (s; i (s;

) = log(1 + s2 ) ) = (1 ? e? s2 )

(2) (3)

f (x) =

We aim towards a probabilistic interpretation and optimal estimates of the model parameters by maximizing the 3

P (P i Pae2A i ae ( ri ae(x)Wi ae ri ae(x)>) + j j af ( rj af (x)Wj af rj af (x)>))

1 2

(7)

Displaced Minima at the Bounds -12080 Real Quadratic Model

-12085

Cost Function

-12090

-12095

-12100

Figure 2: Image Processing Operations: original image, motion boundaries, edge detection, ?ow ?eld in horizontal direction. When evaluating sampled hypothesis, only the edge based term in (7) is used (see section 4.2).

-4

-2 0 2 4 Multiples of Standard Deviation

6

8

Figure 3: Displaced minimum due to bounds constraints For optimization, we use a second order damped Newton trust region method, where a descent direction is chosen by solving the regularized system [12]:

4 Hybrid Optimization

Our search technique uses a combination of robust, consistent, local continuous optimization and more global (discrete) informed sampling.

(H + W ) x = ?g

(11)

4.1 Robust Bound Consistent Continuous Optimization

The model prediction in the image involves a complex generalized transformation including non-linear kinematics and perspective projection. The Jacobian matrix of this trans@T encodes the connection between difformation, J @x ferential quantities in the parameter and observation space, respectively. The robust gradient and Hessian corresponding to the predicted model feature i and the assignments a 2 A can be derived:

=

gi = Ji> Hi Ji>

X

X

a2A

a2A

0 Wia ria ia

where W is a symmetric positive-de?nite matrix and is a dynamically chosen weighting factor. Joint constraints are handled in the optimizer, by projecting the gradient onto the current active constraint set. We ?nd that imposing anatomical constraints on the joints effectively regularizes the tracking, providing an ef?cient way of bringing a weak form of prior knowledge into the problem. Also, notice that adding bound constraints in effect changes the shape of the cost function, and hence the minimum reached. In Figure 3 we plot a 1D slice through the constrained cost function together with a Taylor expansion based on the quadratic approximation. Notice nonzero gradient at minimum due to the presence of the bounds. The gradient changes abruptly because the active-set projection changes the motion direction during the slice to maintain consistency with the constraints.

(8)

4.2 Multiple Hypothesis Algorithm

We represent the evolution of our distribution over time as a set of hypotheses, weighted based on their costs. Although representational schemes based on propagating multiple hypotheses (samples) tend to increase the robustness in the estimated model, the great dif?culty with high-dimensional distributions is hitting their typical sets (areas where most of the probability mass is concentrated). Consequently, informed ways of driving the sampling are needed. The 4

( i 0aWi a + 2 00(Wi a ri a)(Wi a ri a )>)Ji i

(9)

The gradient and Hessian corresponding to all observations are assembled, together with prior contributions:

g = go + 5fp; H = Ho + 52 fp

(10)

main techniques developed in the vision sampling community involve either (i) a combination of careful regularization of the cost function, use of prior motion models, and importance sampling, or (ii) the adoption of a hierarchical search, actually a form of Gibbs sampling (assuming Gaussian marginals) referred as partitioned sampling [21], or (iii) general MCMC sampling techniques based on annealing. Although we acknowledge the importance of both careful cost function design and the incorporation of more speci?c prior knowledge on human motion as a component of a robust system, we observe that no previous sampling technique has attempted to base its search on the fundamental, and therefore general intrinsic component of any generative model, the generalized transform that maps points in model parameter space to observation-associated predictions, thus automatically including non-linear kinematics, perspective projection and data-association uncertainties and singularities. For example, such an algorithm is able to detect problems emphasized by [11] and to focus the search towards those directions in parameter space. We propose a general algorithm in which each hypothesis over model con?gurations is subjected at each timestep to a robust continuous optimization. The estimation uncertainty of each hypothesis is represented by its covariance matrix (inverse Hessian) at its convergence point. An eigenvalue analysis primarily recovers the principal directions and we sample primarily along directions with high uncertainty, by normalizing according to their standard deviations. We generate and evaluate a set of hypotheses which is added to the hypothesis pool of the next frame. For speed, the evaluation uses only the edge component of the cost function, although some information from optical ?ow is automatically included by weighting by the con?dence map derived from the motion boundaries. When all hypotheses have been processed, the pool is pruned by choosing the best k hypotheses which are then propagated to the next time step. We have empirically studied the form of the cost function by searching along uncertain directions for various model con?gurations. The careful selection of image descriptors, combined with the robust continuous optimization usually produces a smooth, singly peaked, distribution over model parameters. Locally multi-modal behavior is found in certain con?gurations as shown in Figure 5, which corresponds to one of the hypotheses at times 0.8sec and 0.9sec in the human tracking sequence of Figure (8). We have also studied the cost function along uncertain directions at much larger scales in parameter space (see Figure 6) and notice that we recover the expected robust shape of the distribution, without too many spurious local minima. Consequently the conjunction of robust cost function design and informed search is likely to be computationally 5

k number of hypothes to propagate is ns number of smallest singular vectors toconsider pf perturbation factor along singular dir ections

MultipleHypothesisA lgorithm(k, ns, pf) { t = 1 hypothesisSet(t) = InitializeModel() for all h ε hypothesisSet(t) c { nextSet = {} (hi,Hi) = RobustContinous Optimize(hc) (v, λ) = GeneralizedEigen Decomposition(H ) i j=1..N for j = N down N ? ns to { σj = 1 / sqrt( j) λ for p = ?p to pf f { nh = hi + pσjvj i Evaluate Cost(nhi) Add(nh , nextSet) i } } hypothesisSet(t+1) = Prune(k, nex tSet) t=t+1 } }

Figure 4: Multiple Hypothesis Algorithm ef?cient. The multiple hypothesis algorithm is brie?y summarized in Figure (4).

5 Model Initialization

The visual tracking starts with a set of initial hypotheses resulting from a model initialization process. Correspondences need to be speci?ed between model joint locations and approximate joint positions of the subject in the initial image. Given this input, we perform a consistent estimation of joint angles and body dimensions. Previous approaches to model initialization from similar input and a single view, haven’t fully addressed the generality and consistency problems [25, 2], not enforcing the joint limits constraints and making assumptions regarding either restricted camera models or restricted human subject poses in the image, respectively. Alternative approaches based on approximate pose recovery based on learned silhouette appearance [6, 23], might be used to bootstrap an algorithm like the one we propose. Our initialization algorithm is a hierarchical three step process that follows a certain parameter estimation scheduling policy. Each stage of the estimation process is essentially based on the formulation described in Section 4.1. Hard joint limits constraints are automatically enforced

Local Minima -11290 Real Quadratic Model

-11300

-11310

Cost Function

-11320

-11330

-11340

-11350

-11360 -10

-8

-6 -4 -2 0 Multiples of Standard Deviation Local Minima

2

4

-11950 Real Quadratic Model

-11955

-11960

-11965

-11970

-11975

-11980

ing parameters on the left and right sides of the body are “mirrored”, while we collect measurements from the entire body (see below). Initialization proceeds as follows. Firstly, we solve an estimation problem under the given 3D model to 2D image correspondences (by minimizing the projection error), and prior intervals on the internal proportions and sizes of the body parts (namely parameters xa , xi and and some of xd ). Secondly we optimize only over the remaining volumetric body sizes alone (limb crossections and their tapering parameters xd ) while holding the other parameters ?xed, using both the given correspondences and the local contour signal from image edges. Finally, we re?ne all model parameters (x) based on similar image information as in the second stage. The covariance matrix corresponding to the ?nal estimate is used to generate an initial set of hypotheses, which are propagated in time using the algorithm described in (4.2). While the process is still heuristic, it gives a balance between stability and ?exibility. In practice we ?nd that enforcing the joint constraints, mirror information and prior bounds on the variation of body parameters gives far more stable and satisfactory results. However, in the monocular case, the initialization always remains uncertain in some directions, a problem that is probably only soluble by fusing pose information from multiple images.

Cost Function

-11985

-11990 -2 0 2 4 6 Multiples of Standard Deviation 8

6 Experiments

The experiments we show consists of an 8sec arm tracking sequence and a 1.2sec full body one in noisy image sequences with complex backgrounds and cluttered subjects as well as signi?cant self-occlusion. Both sequences were shot at 25 FPS interlaced video rate. The experiments were run on a SGI O2 at 270 Mhz using an unoptimized implementation. They take around 10 sec/frame for the hand experiment and around 4 min/frame for the full human body sequence, most of the time being spent evaluating the cost function. Shots showing the model (rendered) overlayed on the image are shown. For the arm sequence, only half is shown while the other half involves a similar reversed trajectory back to the starting position. The hand sequence involves 7 d.o.f which are tracked using just 3 hypothesis resulting from sampling along the smallest uncertain direction, while the full human body sequence involves 30 d.o.f which are tracked using 7 hypothesis based on discretizing the last 3 uncertain directions, each with 2 hypotheses, followed by pruning and propagation in each frame. The motion is successfully tracked over the entire sequence in both cases. Notice that we are tracking in a cluttered background, with specular lighting and loose ?tting clothing. As the 6

Figure 5: Local minima along the direction corresponding to the highest uncertainty, human sequence 0.8s and 0.9s

Cost Function over an Exponentially Sampled Scale -3000 Singv_28 Singv_19 Singv_17 Singv_1

-4000

-5000

-6000

Cost Function

-7000

-8000

-9000

-10000

-11000

-12000 -1000 -500 0 500 1000

Multiples of Standard Deviation

Figure 6: Cost function at larger scales in parameter space

at all stages by the optimization procedure, and correspond-

hand sequence was shot with a camera situated quite close to the subject, the deformations undergone by the arm muscles are signi?cant. The imperfections in our arm model are also apparent. This caused problems when using a tracker based on a single hypothesis and non-robust Gaussian error distributions, especially at the points where the arm edges coincide with the pillar and when the arm self-occludes. Our focus of attention strategy based on motion boundaries helps to disambiguating the dif?cult con?gurations in which the arm (and the full human body), passes the white pillar with strong contrast edges. As emphasized in Section 4.2, we ?nd multi-modal behavior during dif?cult con?gurations when the motions are far from parallel with the image plane, for instance between 2.2sec-4sec in the arm sequence, and over almost the entire full body sequence.

t=0.1s

t=0.75s

7 Conclusions

We have presented a novel multiple hypothesis approach to monocular human motion tracking. The approach is robust at various levels, namely: (1) in the cost function involving the extracted edge and intensity descriptors, and data association using a weighted assignment scheme based on focus of attention; (2) at the optimization level involving joint limit consistent continuous optimization based on realistic likelihood models, and at a more global scale by representing the posterior distribution as a discrete set of hypotheses. An important component of the method is an ef?cient hypothesis generation scheme, which automatically focuses the search in the areas of the parameter space that are poorly estimated due to the combined effect of non-linear kinematics, perspective projection and data-association nonlinearities. We have also introduced a semi-automatic algorithm for pose and body proportion acquisition from a single image and reported experimental results in complex scenes. Future work will be directed towards a more rigorous analysis of the uncertainty during initialization and its propagation in the tracking stage, and to the inclusion and application of general priors on human poses and motions.

t=1.7s

t=2.2s

t=2.8s

t=4s

Acknowledgments

Figure 7: Shots in the arm tracking sequence

References

[1] A.Barr, “Global and local deformations of solid primitives,” Computer Graphics, 18:21-30, 1984. [2] C.Barron and I.Kakadiaris, “Estimating Anthropometry and Pose from a Single Image,” CVPR, pp. 669-676, 2000. [3] M.Black and P.Anandan, “The Robust Estimation of Multiple Motions:Parametric and Piecewise Smooth Flow Fields,” CVIU, Vol.63, No.1, pp.75-104, 1996. [4] M.Black, A.Rangarajan, “On the uni?cation of line processes, outlier rejection, and robust statistics with applications in early vision,” IJCV, Vol. 19, No. 1, pp. 57-92, July, 1996. [5] A.Blake, B.North and M.Isard, “Learning multi-class dynamics”, ANIPS 11, pp.389-395, 1999. [6] M.Brand, “Shadow Puppetry,” ICCV, pp.1237-1244, 1999.

7

t=0.1s

t=0.9s

t=1.1s

t=0.1s

t=0.3s

Figure 9: Images in the human body tracking sequence

Articulated Objects,” ICCV, pp.612-617, 1995. [9] T.Cham and J.Rehg, “A Multiple Hypothesis Approach to Figure Tracking”, CVPR, Vol.2, pp.239-245, 1999. [10] J.Deutscher, B.North, B.Bascle, A.Blake, “Tracking through Singularities and Discontinuities by Random Sampling,” ICCV, pp.11441149, 1999. [11] J.Deutscher, A.Blake, I.Reid, “Articulated Body Motion Capture by Annealed Particle Filtering,” CVPR, 2000. [12] R.Fletcher, “Practical Methods of Optimization,” John Wiley, 1987. [13] D.Gavrila and L.Davis, “3-D Model Based Tracking of Humans in Action:a Multiview Approach,” CVPR, pp. 73-80, 1996. [14] D.Gavrila, “The Visual Analysis of Human Movement:A Survey,” CVIU, Vol.73, No.1, pp.82-98, 1999. [15] L.Gonglaves, E.Bernardo, E.Ursella, P.Perona, “Monocular Tracking of the human arm in 3D”, ICCV, pp.764-770, 1995. [16] N.Howe, M.Leventon, W.Freeman, “Bayesian Reconstruction of 3D Human Motion from Single-Camera Video”, ANIPS, 1999. [17] N.Jojic, M.Turk, T.Huang, “Tracking Self-Occluding Articulated Objects in Dense Disparity Maps,” ICCV, pp.123-130, 1999. [18] I.Kakadiaris and D.Metaxas, “Model-Based Estimation of 3D Human Motion with Occlusion Prediction Based on Active MultiViewpoint Selection,” CVPR, pp. 81-87, 1996. [19] J.Kuch and T.Huang “Vision-based modeling and tracking for virtual teleconferencing and telecollaboration,” ICCV, pp.666-671, 1995. [20] M.Leventon and W.Freeman, “Bayesian Estimation of 3-d Human Motion from an Image Sequence,”, TR-98-06, MERL, 1998. [21] J.MacCormick and M.Isard, “Partitioned sampling, articulated objects, and interface-quality hand tracker”, ECCV, Vol.2, pp.3-19, 2000. [22] D.Morris and J.Rehg, “Singularity Analysis for Articulated Object Tracking”, , pp. 289-296, CVPR 1998. [23] R.Rosales and S.Sclaroff, “Inferring Body Pose without Tracking Body Parts,” CVPR, pp.721-727, 2000. [24] H.Sidenbladh, M.Black, D.Fleet, “Stochastic Tracking of 3D Human Figures Using 2D Image Motion,” ECCV, 2000. [25] C.J.Taylor, “Reconstruction of Articulated Objects from Point Correspondences in a Single Uncalibrated Image,” CVPR, pp.677-684, 2000. [26] B.Triggs, P.McLauchlan, R.Hartley, A.Fitzgibbon, “Bundle Adjustment - A Modern Synthesis,” Vision Algorithms: Theory and Practice, Springer-Verlag, LNCS 1883, 2000.

t=0.6s

t=0.9s

t=1.1s

t=1.2s

Figure 8: Shots during human body tracking sequence

[7] C.Bregler and J.Malik, “Tracking People with Twists and Exponential Maps,” CVPR, 1998. [8] J.Rehg and T.Kanade “Model-Based Tracking of Self Occluding

8

[27] S.Wachter and H.Nagel, “Tracking Persons in Monocular Image Sequences,” CVIU, 74(3):174-192, 1999. [28] C.Wren and A.Pentland, ”DYNAMAN; A Recursive Model of Human Motion,” MIT Media Lab Tech. Report, No. 451.

9

赞助商链接

- 3D reconstruction of human skeleton from single images or monocular video sequences
- Human Motion Reconstruction and Animation from Video Sequences
- MCMC-based particle filtering for tracking a variable number of interacting targets
- Maintaining multi-modality through mixture tracking
- Using particles to track varying numbers of interacting people
- A Multiple Hypothesis Approach to Figure Tracking
- A Robust Multiple Hypothesis Approach
- Feature Tracking and Motion Factorization from Monocular Video
- Robust Tracking of Human Motion
- Robust visual tracking for multiple targets
- Skeleton-Based Motion Capture for Robust Reconstruction of Human Motion
- ROBUST TRACKING CONTROLLER DESIGN WITH APPLICATION TO THE MOTION CONTROL OF AN X-Y FEED TAB
- Robust Tracking of Athletes Using Multiple Features of Multiple Views
- Multiple-model probability hypothesis density filter for tracking maneuvering targets
- Robust Real-Time Multiple Target Tracking

更多相关文章：
更多相关标签：