The task of classifying sequences of data is a key technology in a broad range of applications. In econometrics and forecasting applications, classification of data sequences which can represent the inventory of product in a distribution channel over time, or the price of a stock or other financial instrument over time can be crucial. Different classes may, for instance, signify various important stock market states. In human motion applications, data sequences which represent the pose of the human body or one of its parts over time can be classified into categories. For example, a running motion can be distinguished from walking. Movements of hands can be automatically interpreted as individual signs in American Sign Language and then used to interface to a computer.
In classification applications, the goal is to assign class labels to observed data sequences or trajectories. Dynamic models can be useful in classification because they provide an efficient coding of the set of all possible trajectories. For example, suppose that two distinct gestures can be described with separate linear dynamical models. For each model, it is straightforward to compute an error signal, called the innovation, which measures the extent to which the model predicts the observations. This innovation can be used directly for classification. The parameters of the linear dynamic models therefore provide a very compact representation of the class of trajectories that make up a particular gesture.
Sets of dynamic models can be used to model qualitatively different regimes of a trajectory associated with one temporal event. For instance, a hand gesture can be segmented into three motion regimes or phases: preparation, stroke and retraction. Each regime can then be associated with a different linear model. The sequence of regimes can be governed by a model switching process.
Switching linear dynamic system (SLDS) models consist of a set of linear dynamic models and a switching variable that determines which model is in effect at any given point in time. In addition, fully connected SLDS models assume that there are temporal dependencies between the switching variables as well as the states of different linear dynamic models. SLDS models are attractive because they can describe complex dynamics using simple linear models as building blocks. Given an SLDS model, the inference problem is to estimate the sequence of model states that best explains an input data sequence. Unfortunately, exact inference in SLDS models is computationally intractable, due to the large number of possible combinations of linear models over time.
One prior method for inference using fully connected SLDS models is described in “Estimation and Tracking: Principles, Techniques, and Software” by Bar-Shalom et al., Artech House, Inc. 1993. In this method, approximate inference is achieved by truncating or collapsing the number of discrete components in the evolving model. The models were used to detect different motion regimes while tracking a maneuvering target.
In another prior method the switching variable determines which linear model is coupled to the measurement at each time instant. See Ghahramani et al., “Variational Learning for Switching State-Space Models” which will appear in the journal Neural Computation. This method can produce decoupled linear models which reach steady-state before the data series is adequately modeled. It stands in contrast to the prior methods with fully coupled SLDS in which all models are coupled through a single state space. The method was used to segment regimes of no breathing and gasping breathing in data collected from patients with sleep apnea.
A different prior method in “Time-series Classification Using Mixed-State Dynamic Bayesian Networks,” by Pavlovic et al., Proc. of Computer Vision and Pattern Recognition, pages 609-615, June, 1999, considers a single linear dynamical model whose input is modeled as a discrete Markov process. This model explains all measurement variability as a consequence of the changes in input, which may not be true in general. The model was applied to classification of computer mouse-drawn symbols.
Another prior method proposes particle filters as an alternative to using linear models as the building blocks in a switching framework. See Blake et al., “Learning Multi-Class Dynamics,” Advances in Neural Information Processing Systems (NIPS '98), pages 389-395, 1998. The use of a nonparametric, particle-based model can be inefficient in domains where linear models are a powerful building block. Nonparametric methods are particularly expensive when applied to large state spaces, since they are exponential in the state space dimension.
In another prior method, a Hidden Markov Model with an entropic prior is proposed for dynamics learning from sparse input data. See Brand, “Pattern discovery via entropy minimization,” Technical Report TR98-21, Mitsubishi Electric Research Lab, 1998. The method is applied to the synthesis of facial animation and, to a certain extent, the segmentation of facial expressions from voice data, e.g., Brand, “Voice puppetry,” Proceeding of SIGGRAPH99, 1999. The dynamic models produced by this method are time invariant. Each state space neighborhood has a unique distribution over state transitions. In addition, the use of entropic priors results in fairly deterministic models learned from a moderate corpus of training data. In many applications, time-invariant models are unlikely to succeed, since different state space trajectories can originate from the same starting point, depending upon the class of motion being performed.
Modeling Human Dynamics
Technologies for analyzing the motion of the human figure play a key role in a broad range of applications, including computer graphics, user-interfaces, surveillance, and video editing. A motion of the figure can be represented as a trajectory in a state space which is defined by the kinematic degrees of freedom of the figure. Each point in state space represents a single configuration or pose of the figure. A motion such as a plié in ballet is described by a trajectory along which the joint angles of the legs and arms change continuously.
A key issue in human motion analysis and synthesis is modeling the dynamics of the figure. While the kinematics of the figure define the state space, the dynamics define which state trajectories are possible (or probable) in that state space. Prior methods for representing human dynamics have been based on analytic dynamic models. Analytic models are specified by a human designer. They are typically second order differential equations relating joint torque, mass, and acceleration.
The field of biomechanics is a source of complex and realistic analytic models of human dynamics. From the biomechanics point of view, the dynamics of the figure are the result of its mass distribution, joint torques produced by the motor control system, and reaction forces resulting from contact with the environment, e.g., the floor. Research efforts in biomechanics, rehabilitation, and sports medicine have resulted in complex, specialized models of human motion. For example, detailed walking models are described in Inman et al., “Human Walking,” Williams and Wilkins, 1981.
The biomechanical approach has two drawbacks. First, the dynamics of the figure are quite complex, involving a large number of masses and applied torques, along with reaction forces which are difficult to measure. In principle, all of these factors must be modeled or estimated in order to produce physically valid dynamics. Second, in some applications, we may only be interested in a small set of motions, such as a vocabulary of gestures. In the biomechanical approach it may be difficult to reduce the complexity of the model to exploit this restricted focus. Nonetheless, biomechanical models have been applied to human motion analysis.
A prior method for visual tracking uses a biomechanically-derived dynamic model of the upper body. See Wren et al., “Dynamic models of human motion,” Proceeding of the Third International Conference on Automatic Face and Gesture Recognition, pages 22-27, Nara, Japan, 1998. The unknown joint torques are estimated along with the state of the arms and head in an input estimation framework. A Hidden Markov Model is trained to represent plausible sequences of input torques. This prior art does not address the problem of modeling reaction forces between the figure and its environment. An example is the reaction force exerted by the floor on the soles of the feet during walking or running.
Therefore, there is a need for classification methods for fully coupled SLDS models that can segment data sequences into regimes.
Technologies for analyzing the motion of the human figure play a key role in a broad range of applications, including computer graphics, user-interfaces, surveillance, and video editing. A motion of the figure can be represented as a trajectory in a state space which is defined by the kinematic degrees of freedom of the figure. Each point in state space represents a single configuration or pose of the figure. A motion such as a plié in ballet is described by a trajectory along which the joint angles of the legs and arms change continuously.
A key issue in human motion analysis and synthesis is modeling the dynamics of the figure. While the kinematics of the figure define the state space, the dynamics define which state trajectories are possible (or probable) in that state space. Prior methods for representing human dynamics have been based on analytic dynamic models. Analytic models are specified by a human designer. They are typically second order differential equations relating joint torque, mass, and acceleration.
The field of biomechanics is a source of complex and realistic analytic models of human dynamics. From the biomechanics point of view, the dynamics of the figure are the result of its mass distribution, joint torques produced by the motor control system, and reaction forces resulting from contact with the environment (e.g. the floor). Research efforts in biomechanics, rehabilitation, and sports medicine have resulted in complex, specialized models of human motion. For example, detailed walking models are described in Inman et al., “Human Walking,” Williams and Wilkins, 1981.
The biomechanical approach has two drawbacks. First, the dynamics of the figure are quite complex, involving a large number of masses and applied torques, along with reaction forces which are difficult to measure. In principle all of these factors must be modeled or estimated in order to produce physically-valid dynamics. Second, in some applications we may only be interested in a small set of motions, such as a vocabulary of gestures. In the biomechanical approach it may be difficult to reduce the complexity of the model to exploit this restricted focus. Nonetheless, biomechanical models have been applied to human motion analysis.
A prior method for visual tracking uses a biomechanically-derived dynamic model of the upper body. See Wren et al., “Dynamic models of human motion,” Proceeding of the Third International Conference on Automatic Face and Gesture Recognition, pages 22-27, Nara, Japan, 1998. The unknown joint torques are estimated along with the state of the arms and head in an input estimation framework. A Hidden Markov Model is trained to represent plausible sequences of input torques. This prior art does not address the problem of modeling reaction forces between the figure and its environment. An example is the reaction force exerted by the floor on the soles of the feet during walking or running.
We present a method for classification of data sequences modeled as fully coupled switching linear dynamic models (SLDSs). The method uses approximate Viterbi inference to find a most likely sequnece of switching states.
The classification method can be used to classify motion regimes learned from figure motion data corresponding to classes of human activity such as running, walking, etc. Inference produces a single sequence of switching modes which best describes a motion trajectory in the figure state space. This sequence segments the figure motion trajectory into motion regimes learned from data.
Described herein is a new class of approximate learning methods for switching linear dynamic (SLDS) models. These models consist of a set of linear dynamic system (LDS) models and a switching variable that indexes the active model. This new class has three advantages over dynamics learning methods known in the prior art:
In addition, this method can be applied to the problem of learning dynamic models for human motion from data. It has three advantages over analytic dynamic models known in the prior art:
A wide range of learning algorithms can be cast in the framework of Dynamic Bayesian Networks (DBNs). DBNs generalize two well-known signal modeling tools: Kalman filters for continuous state linear dynamic systems (LDS) and Hidden Markov Models (HMMs) for discrete state sequences. Kalman filters are described in Anderson et al., “Optimal Filtering,” Prentice-Hall, Inc., Englewood Cliffs, N.J., 1979. Hidden Markov Models are reviewed in Jelinek, “Statistical methods for speech recognition,” MIT Press, Cambridge, Mass., 1998.
Dynamic models learned from sequences of training data can be used to predict future values in a new sequence given the current values. They can be used to synthesize new data sequences that have the same characteristics as the training data. They can be used to classify sequences into different types, depending upon the conditions that produced the data.
We focus on a subclass of DBN models called Switching Linear Dynamics Systems. Intuitively, these models attempt to describe a complex nonlinear dynamic system with a succession of linear models that are indexed by a switching variable. The switching approach has an appealing simplicity and is naturally suited to the case where the dynamics are time-varying.
We present a method for approximate inference in fully coupled switching linear dynamic models (SLDSs). Exponentially hard exact inference is replaced with approximate inference of reduced complexity.
The first preferred embodiment uses Viterbi inference jointly in the switching and linear dynamic system states.
The second preferred embodiment uses variational inference jointly in the switching and linear dynamic system states.
Parameters of a fully connected SLDS model are learned from data. Model parameters are estimated using a generalized expectation-maximization (EM) algorithm. Exact expectation/inference (E) step is replaced with one of the three approximate inference embodiments.
The learning method can be used to model the dynamics of human motion. The joint angles of the limbs and pose of the torso are represented as state variables in a switching linear dynamic model. The switching variable identifies a distinct motion regime within a particular type of human motion.
Motion regimes learned from figure motion data correspond to classes of human activity such as running, walking, etc. Inference produces a single sequence of switching modes which best describes a motion trajectory in the figure state space. This sequence segments the figure motion trajectory into motion regimes learned from data.
Accordingly, a method for classifying portions of an input measurement sequence into a plurality of regimes includes associating each of a plurality of dynamic models with one a switching state such that a model is selected when its associated switching state is true. A state transition record is determined by determining and recording, for a given measurement of the sequence and for each switching state, an optimal prior switching state, based on the input sequence, where the optimal prior switching state optimizes a transition probability. An optimal final switching state is determined for a final measurement. A switching state sequence is determined by backtracking through the state transition record from the optimal final switching state. Finally, portions of the input sequence are classified into different regimes, responsive to the switching state sequence.
In one embodiment of the present invention, classifying depends upon conditions existing at the time the sequence was created.
Regimes can be, for example, motion regimes, and in particular, human motion. Human motion includes, but is not limited to, walking, jogging, running, jumping, sitting, and climbing, and ascending and descending a staircase.
An embodiment of the present invention is capable of identifying an individual based on observed dynamics of the individual's motion in an image sequence. This can be useful, for example, in identifying a criminal suspect from the image sequence.
Another embodiment classifies sequences into motions to conduct surveillance. For example, certain activities, such as opening a door or dropping a package, can be identified by classification.
In one embodiment, one or more constraints can be imposed on the classification.
Each motion can be an individual sign of a sign language In yet another embodiment, classification of a motion serves as input to a computer user interface.
In yet another embodiment, sets of dynamic models are used to model qualitatively different regimes of a trajectory with a single temporal event.
Yet another embodiment selects key frames from an input sequence in response to the classification, and performs video compression by transmitting key frames at a low sampling rate.
While the above embodiments are based on Viterbi techniques, other embodiments of the present invention are based on variational techniques. For example, a method for classifying portions of an input sequence of measurements into a plurality of regimes, given a set of possible switching states, includes associating each of a plurality of dynamic models with a switching state such that a dynamic model is selected when its associated switching state is true, where the switching state at a particular instance is determined by a switching model. The dynamic model is then decoupled from the switching model. Parameters of the decoupled dynamic model are determined responsive to a switching state probability estimate. A state of the decoupled dynamic model corresponding to a measurement at the particular instance is estimated, responsive to the input sequence. Parameters of the decoupled switching model are then determined responsive to the dynamic state estimate. A probability is estimated for each possible switching state of the decoupled switching model. A switching state sequence is determined based on the estimated switching state probabilities. Finally, portions of the input sequence are classified into different regimes, responsive to the determined switching state sequence.
The foregoing and other objects, features and advantages of the invention will be apparent from the following more particular description of preferred embodiments of the invention, as illustrated in the accompanying drawings in which like reference characters refer to the same parts throughout the different views. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating the principles of the invention.
Switching Linear Dynamic System Model
The system can be described using the following set of state-space equations:
xt+1=A(st+1)xt+vt+1(st+1),
yt=Cxt+wt, and
x0=v0(s0) (Eq. 1)
for the physical system 10, and
Pr(st+1|st)=st+1′Πst, and
Pr(s0)=π0
for the switching model 12.
Here, xtεN denotes the hidden state of the LDS 10 at time t, and vt is the state noise process. Similarly, ytε
M is the observed measurement at time t, and wt is the measurement noise. Parameters A and C are the typical LDS parameters: the state transition matrix 14 and the observation matrix 16, respectively. Assuming the LDS models a Gauss-Markov process, the noise processes are independently distributed Gaussian:
vt(st)˜N(0, Q(st)), t>0
v0(s0)˜N(x0(s0), Q0(s0))
wt˜N(0, R).
where Q is the state noise variance and R is the measurement noise variance. The notation s′ is used to indicate the transpose of vector s.
The switching model 12 is assumed to be a discrete first-order Markov process. State variables of this model are written as st. They belong to the set of S discrete symbols {e0, . . . , eS−1}, where ei is, for example, the unit vector of dimension S with a non-zero element in the i-th position. The switching model 12 is a first-order discrete Markov model defined with the state transition matrix Π whose elements are
Π(i,j)=Pr(St+1)=ei|st=ej), (Eq. 2)
and which is given an initial state distribution Π0.
Coupling between the LDS and the switching process is full and stems from the dependency of the LDS parameters A and Q on the switching process state st. Namely,
A(st=ei)=Ai
Q(st=ei)=Qi
In other words, switching state st determines which of S possible models {(A0, Q0), . . . , (AS−1, QS−1)} is used at time t.
where YT, XT, and ST denote the sequences, of length T, of observation and hidden state variables, and switching variables, respectively. For example, YT={y0, . . . , yT−1}. In this dependency graph 20, the coupling between the switching states and the LDS states is full, i.e., switching states are time-dependent, LDS states are time-dependent, and switching and LDS states are intradependent.
We can now write an equivalent representation of the fully coupled SLDS as the above DBN, assuming that the necessary conditional probability distribution functions (pdfs) are appropriately defined. Assuming a Gauss-Markov process on the LDS, it follows:
xt+1|xt, st+1=ei˜N(Aixt, Qi),
yt|xt˜N(Cxt, R),
x0|s0=ei˜N(x0,i, Q0, i)
Recalling the Markov switching model assumption, the joint pdf of the complex DBN of duration T, or, equivalently, its Hamiltonian, where the Hamiltonian H(x) of a distribution P(x) is defined as any positive function such that P(x)=[(exp(−H(x)))/(ΣΨexp(−H(Ψ)))], can be written as:
Inference
The goal of inference in SLDSs is to estimate the posterior probability of the hidden states of the system (st and xt) given some known sequence of observations YT and the known model parameters, i.e., the likelihood of a sequence of models as well as the estimates of states. Namely, we need to find the posterior
P(XT, ST|YT)=Pr(XT, ST|YT),
or, equivalently, its “sufficient statistics”. Given the form of P it is easy to show that these are the first and second-order statistics: mean and covariance among hidden states xt, xt−1, st, st−1.
If there were no switching dynamics, the inference would be straightforward—we could infer XT from YT using an LDS inference approach such as smoothing, as described by Rauch, “Solutions to the linear smoothing problem,” IEEE Trans. Automatic Control, AC-8(4):371-372, October 1963. However, the presence of switching dynamics embedded in matrix P makes exact inference more complicated. To see that, assume that the initial distribution of x0 at t=0 is Gaussian. At t=1, the pdf of the physical system state x1 becomes a mixture of S Gaussian pdfs since we need to marginalize over S possible but unknown models. At time t, we will have a mixture of St Gaussians, which is clearly intractable for even moderate sequence lengths. It is therefore necessary to explore approximate inference techniques that will result in a tractable learning method. What follows are three preferred embodiments of the inference step.
Approximate Viterbi Inference Embodiment
The task of the Viterbi approximation approach of the present invention is to find the most likely sequence of switching states st for a given observation sequence YT. If the best sequence of switching states is denoted S*T, then the desired posterior P(XT, ST|YT) can be approximated as
P(XT, ST|YT)=P(XT|ST, YT) P(ST|YT)≈P(XT|ST, YT) δ(ST−S*T), (Eq 4)
where δ(x)=1 if x=0 and δ(x)=0 if x≠0. In other words, the switching sequence posterior P(ST|YT) is approximated by its mode. Applying Viterbi inference to two simpler classes of models, discrete state hidden Markov models and continuous state Gauss-Markov models is well-known. An embodiment of the present invention utilizes an algorithm for approximate Viterbi inference that generalizes the two approaches.
We would like to determine the switching sequence S*T such that S*T=arg maxSTP(ST|YT). First, define the following probability Jt,i up to time t of the switching state sequence being in state i at time t given the measurement sequence Yt:
If this quantity is known at time T, the probability of the most likely switching sequence S*T is simply P(S*T|YT)∝maxiJT−1,i. In fact, a recursive procedure can be used to obtain the desired quantity. To see that, express Jt,i in terms of Js at t−1. It follows that
where we denote
Jt|t−1,i,j=P(yt|st=ei,st−1=ej,S*t−2(j), Yt−1) P(st=ei|st−1=ej) (Eq. 7)
as the “transition probability” from state j at time t−1 to state i at time t. Since this analysis takes place relative to time t, we refer to st as the “switching state,” and to st−1 as the “previous switching state.” S*t−2(i) is the “best” switching sequence up to time t−1 when SLDS is in state i at time t−1:
Hence, the switching sequence posterior at time t can be recursively computed from the same at time t−1. The two scaling components in Jt|t−1,i,j are the likelihood associated with the transition j→i from t−1 to t, and the probability of discrete SLDS switching from j to i.
To find the likelihood term, note that concurrently with the recursion of Equation 6, for each pair of consecutive switching states j,i at times t−1,t, one can obtain the following statistics using the Kalman filter:
where {circumflex over (x)}t|t,i is the “best” filtered LDS state estimate at t when the switch is in state i at time t and a sequence of t measurements, Yt, has been processed; {circumflex over (x)}t|t−1,i,j and {circumflex over (x)}t|t,i,j are the one-step predicted LDS state and the “best” filtered state estimates at time t, respectively, given that the switch is in state i at time t and in state j at time t−1 and only t−1 measurements are known. The two sets {{circumflex over (x)}t|t−1,i,j} and {{circumflex over (x)}t|t,i,j}, where i and j take on all possible values, are examples of “sets of continuous state estimates.” The set {{circumflex over (x)}t|t−1,i,j} is obtained through “Viterbi prediction.” Similar definitions are easily obtained for filtered and predicted state variance estimates, Σt|t,i and Σt|t−1,i,j respectively. For a given switch state transitions j→i it is now easy to establish relationship between the filtered and the predicted estimates. From the theory of Kalman estimation, it follows that for transitions j→i the following time updates hold:
{circumflex over (x)}t|t−1,i,j=Ai{circumflex over (x)}t−1|t−1,j (Eq. 9)
Σt|t−1,i,j=AiΣt−1|t−1,jAi′+Qi (Eq. 10)
Given a new observation yt at time t, each of these predicted estimates can now be filtered using a Kalman “measurement update” framework. For instance, the state estimate measurement update equation yields
{circumflex over (x)}t|t,i,j={circumflex over (x)}t|t−1,i,j+Ki,j(yt−C{circumflex over (x)}t|t−1,i,j). (Eq. 11)
where Ki,j is the Kalman gain matrix associated with the transitions j→i.
Appropriate equations can be obtained that link
The likelihood term can then be easily computed as the probability of innovation yt−C{circumflex over (x)}t|t−1,i,j of j→i transition,
yt|st=ei,st−1=ei,S*t−2(j)˜N(yt; C{circumflex over (x)}t|t−1,i,j, CΣt|t−1,i,jC′+R) (Eq. 12)
Obviously, for every current switching state i there are S possible previous switching states. To maximize the overall probability at every time step and for every switching state, one “best” previous state j is selected:
ψt−1,i=arg maxj{Jt|t−1,i,jJt−1,j} (Eq. 13)
Since the best state is selected based on the continuous state predictions from the previous time step, ψt−1,i is referred to as the “optimal prior switching state.” The index of this state is kept in the state transition record entry ψt−1,i. Consequently, we now obtain a set of S best filtered LDS states and variances at time t:
{circumflex over (x)}t|t,i={circumflex over (x)}t|t,i,ψ
Once all T observations YT−1 have been fused to decode the “best” switching state sequence, one uses the index of the best final state,
and then traces back through the state transition record ψt−1,i, setting i*t=Ψt,t*
for i=0, . . . , S−1.
Approximate Variational Inference Embodiment
At step 42, the switching dynamics of one or more SLDS motion models are learned from a corpus of state space motion examples 40. Parameters 46 of each SLDS are re-estimated, at step 44, iteratively so as to minimize the modeling cost of current state sequence estimates that are obtained using the approximate variational inference procedure. Approximate variational inference is developed as an alternative to computationally expensive exact state sequence estimation.
A general structured variational inference technique for Bayesian networks is described in Jordan at al., “An Introduction to Variational Methods For Graphical Models,” Learning In Graphical Models, Kluwer Academic Publishers, 1998. They consider a parameterized distribution Q which is in some sense close to the desired conditional distribution P, but which is easier to compute. Q can then be employed as an approximation of P,
P(XT, ST|YT)≈Q(XT, ST|YT)
Namely, for a given set of observations YT, a distribution Q(XT, ST|η, YT with an additional set of variational parameters h is defined such that Kullback-Leibler divergence between Q(XT, ST|η, YT and p(XT, ST|YT) is minimized with respect to h:
The dependency structure of Q is chosen such that it closely resembles the dependency structure of the original distribution P. However, unlike P, the dependency structure of Q must allow a computationally efficient inference. In our case, we define Q by decoupling the switching and LDS portions of SLDS as shown in FIG. 6.
The two subgraphs 50, 52 are “decoupled,” thus allowing for independent inference, Q(XT, ST|η, YT)=Q(XT|η, YT)QS(ST|η). This is also reflected in the sufficient statistics of the posterior defined by the approximating network, e.g., <xtx′tst>=<xtx′t><st>.
The optimal values of the variational parameters h are obtained by setting the derivative of the KL-divergence with respect to h to zero. We can then arrive at the following optimal variational parameters:
To obtain the expectation terms <St>=Pr(st|q0, . . . , qT−1), we use the inference in the HMM with output “probabilities” qt, as described in Rabiner and Juang, “Fundamentals of Speech Recognition,” Prentice Hall, 1993. Similarly, to obtain <xt>=E[xt|YT], LDS inference is performed in the decoupled time-varying LDS via RTS smoothing. Since Ât, {circumflex over (Q)}t in the decoupled LDS Qx depend on <st> from the decoupled HMM QS, and qt depends on <xt>, <xtx′t>, <xtx′t−1> from the decoupled LDS, Equations 15 and 16 together with the inference solutions in the decoupled models form a set of fixed-point equations. Solution of this fixed-point set yields a tractable approximation to the intractable inference of the original fully coupled SLDS.
Variational parameters in Equations 15 and 16 have an intuitive interpretation. Variational parameters of the decoupled LDS {circumflex over (Q)}t and Ât in Equation 15 define a best unimodal (non-switching) representation of the corresponding switching system and are, approximately, averages of the original parameters weighted by a best estimates of the switching states P(s). Variational parameters of the decoupled HMM log qt(i) in Equation 16 measure the agreement of each individual LDS with the data.
General Pseudo Bayesian Inference Embodiment
The Generalized Pseudo Bayesian (GPB) approximation scheme first introduced in Bar-Shalom et al., “Estimation and Tracking: Principles, Techniques, and Software,” Artech House, Inc. 1993 and in Kim, “Dynamic Linear Models With Markov-Switching,” Journal of Econometrics, volume 60, pages 1-22, 1994, is based on the idea of “collapsing”, i.e., representing a mixture of Mt Gaussians with a mixture of Mr Gaussians, where r<t. A detailed review of a family of similar schemes appears in Kevin P. Murphy, “Switching Kalman Filters,” Technical Report 98-10, Compaq Cambridge Research Lab, 1998. We develop a SLDS inference scheme jointly in the switching and linear dynamic system states, derived from filtering GPB2 approach of Bar-Shalom et al. The algorithm maintains a mixture of M2 Gaussians over all times.
GPB2 is closely related to the Viterbi approximation described previously. It differs in that instead of picking the most likely previous switching state j at every time step t and switching state i, we collapse the M Gaussians (one for each possible value of j) into a single Gaussian.
Consider the filtered and predicted means {circumflex over (x)}t|t,i,j and {circumflex over (x)}t|t−1,i,j, and their associated covariances, which were defined previously with respect to the approximate Viterbi embodiment. Assume that for each switching state i and pairs of states (i,j) the following distributions are defined at each time step:
Pr(st=i|Yt)
Pr(st=i,st−1=j|Yt)
Regular Kalman filtering update can be used to fuse the new measurement yt and obtain S2 new SLDS states at t for each S states at time t−1, in a similar fashion to the Viterbi approximation embodiment discussed previously.
Unlike Viterbi approximation which picks one best switching transition for each state i at time t, GPB2 “averages” over S possible transitions from t−1. Namely, it is easy to see that
Pr(st=i,st−1=j|Yt)≈Pr(yt|{circumflex over (x)}t,i,j) Π(i, j) Pr(st−1=j|Yt−1)
From there it follows immediately that the current distribution over the switching states is
and that each previous state j now has the following posterior
This posterior is important because it allows one to “collapse” or “average” the S transitions into each state i into one average state, e.g.
Analogous expressions can be obtained for the variances St|t,i and St,t−1|t,i.
Smoothing in GPB2 is a more involved process that includes several additional approximations. Details can be found in Kevin P. Murphy, “Switching Kalman Filters,” Technical Report 98-10, Compaq Cambridge Research Lab, 1998. First an assumption is made that decouples the switching model from the LDS when smoothing the switching states. Smoothed switching states are obtained directly from Pr(st|Yt) estimates, namely Pr(st=i|st+1=k, YT)≈Pr(st=i|st+1=k, Yt). Additionally, it is assumed that {circumflex over (x)}t+1|T,i,k≈{circumflex over (x)}t+1|T,k. These two assumptions lead to a set of smoothing equations for each transition (i, k) from t to t+1 that obey an RTS smoother, followed by collapsing similar to the filtering step.
The inference process of the GPB2 embodiment is clearly more involved than those of the Viterbi or the variational approximation embodiments. Unlike Viterbi, GPB2 provides soft estimates of switching states at each time t. Like Viterbi, GPB2 is a local approximation scheme and as such does not guarantee global optimality inherent in the variational approximation. However, Xavier Boyen and Daphne Koller, “Tractable inference for complex stochastic processes,” Uncertainty in Artificial Intelligence, pages 33-42, Madison, Wis., 1998, provides complex conditions for a similar type of approximation in general DBNs that lead to globally optimal solutions.
Learning of SLDS Parameters
Learning in SLDSs can be formulated as the problem of maximum likelihood learning in general Bayesian networks. Hence, a generalized Expectation-Maximization (EM) algorithm can be used to find optimal values of SLDS parameters {A0, . . . , As−1, C, Q0, . . . , Qs−1, R, Π, π0}. A description of generalized EM can be found in, for example, Neal et al., “A new view of the EM algorithm that justifies incremental and other variants,” in the collection “Learning in graphical models,” (M. Jordan, editor), pages 355-368. MIT Press, 1999.
The EM algorithm consists of two steps, E and M, which are interleaved in an iterative fashion until convergence. The essential step is the expectation (E) step. This step is also known as the inference problem. The inference problem was considered in the two preferred embodiments of the method.
Given the sufficient statistics from the inference phase, it is easy to obtain parameter update equations for the maximization (M) step of the EM algorithm. Updated values of the model parameters are easily shown to be
The operator <·> denotes conditional expectation with respect to the posterior distribution, e.g. <xt>=Σs∫xxtP(X, S|Y).
All the variable statistics are evaluated before updating any parameters. Notice that the above equations represent a generalization of the parameter update equations of classical (non-switching) LDS models.
The coupled E and M steps are the key components of the SLDS learning method. While the M step is the same for all preferred embodiments of the method, the E-step will vary with the approximate inference method.
Applications of SLDS
We applied our SLDS framework to the analysis of two categories of fronto-parallel motion: walking and jogging. Fronto-parallel motions exhibit interesting dynamics and are free from the difficulties of 3-D reconstruction. Experiments can be conducted easily using a single video source, while self-occlusions and cluttered backgrounds make the tracking problem non-trivial.
The kinematics of the figure are represented by a 2-D Scaled Prismatic Model (SPM). This model is described in Morris et al., “Singularity analysis for articulated object tracking,” Proceedings of Computer Vision and Pattern Recognition, pages 289-296, Santa Barbara, Calif., June 23-25, 1998. The SPM lies in the image plane, with each link having one degree of freedom (DOF) in rotation and another DOF in length. A chain of SPM transforms can model the image displacement and foreshortening effects produced by 3-D rigid links. The appearance of each link in the image is described by a template of pixels which is manually initialized and deformed by the link's DOFs.
In our figure tracking experiments, we analyzed the motion of the legs, torso, and head, while ignoring the arms. Our kinematic model had eight DOFs, corresponding to rotations at the knees, hip and neck.
Learning
The first task we addressed was learning an SLDS model for walking and running. The training set consisted of eighteen sequences of six individuals jogging and walking at a moderate pace. Each sequence was approximately fifty frames in duration. The training data consisted of the joint angle states of the SPM in each image frame, which were obtained manually.
The two motion types were each modeled as multi-state SLDSs and then combined into a single complex SLDS. The measurement matrix in all cases was assumed to be identity, C=I. Initial state segmentation within each motion type was obtained using unsupervised clustering in a state space of some simple dynamics model, e.g., a constant velocity model. Parameters of the model (A, Q, R, x0, P, π0) were then reestimated using the EM-learning framework with approximate Viterbi inference. This yielded refined segmentation of switching states within each of the models.
Classification
The task of classification is to segment a state space trajectory into a sequence of motion regimes, according to learned models. For instance, in gesture recognition, one can automatically label portions of a long hand trajectory as some predefined, meaningful gestures. The SLDS framework is particularly suitable for motion trajectory classification.
The classification framework follows directly from the framework for approximate Viterbi inference in SLDSs, described previously. Namely, the approximate Viterbi inference 74 yields a sequence of switching states ST (regime indexes) 70 that best describes the observed state space trajectory 70, assuming some current SLDS model parameters. If those model parameters are learned on a corpus of representative motion data, applying the approximate Viterbi inference on a state trajectory from the same family of motions will then result in its segmentation into the learned motion regimes.
Additional “constraints” 72 can be imposed on classification. Such constraints 72 can model expert knowledge about the domain of trajectories which are to be classified, such as domain grammars, which may not have been present during SLDS model learning. For instance, we may know that out of N available SLDS motion models, only M<N are present in the unclassified data. We may also know that those M models can only occur in some particular, e.g., deterministically or stochastically, known order. These classification constraints can then be superimposed on the SLDS motion model parameters in the approximate Viterbi inference to force the classification to adhere to them. Hence, a number of natural language modeling techniques from human speech recognition, for example, as discussed in Jelinek, “Statistical methods for speech recognition,” MIT Press, 1998, can be mapped directly into the classification SLDS domain.
Classification can also be performed using variational and GPB2 inference techniques. Unlike with Viterbi inference, these embodiments yield a “soft” classification, i.e., each switching state at every time instance can be active with some potentially non-zero probability. A soft switching state at time t is
To explore the classification ability of our learned model, we next considered segmentation of sequences of complex motion, i.e., motion consisting of alternations of “jog” and “walk.” Test sequences were constructed by concatenating in random order randomly selected and noise-corrupted training sequences. Transitions between sequences were smoothed using B-spline smoothing. Identification of two motion “regimes” was conducted using the proposed framework in addition to a simple HMM-based segmentation. Multiple choices of linear dynamic system and “jog”/“walk” switching orders were also compared.
Because “jog” and “walk” SLDS models were learned individually, whereas the test data contained mixed “jog” and “walk” regimes, it was necessary to impose classification constraints to combine the two models into a single “jog+walk” model. Since no preference was known a priori for either of the two regimes, the two were assumed equally likely but complementary, and their individual two-state switching models Pjog and Pwalk were concatenated into a single four-state switching model Pjog+walk. Each state of this new model simply carried over the LDS parameters of the individual “jog” and “walk” models.
Estimates of “best” switching states <st> indicated which of the two models can be considered to be driving the corresponding motion segment.
Similar results were obtained with four switching states each and second order SLDS.
Classification experiments on a set of 20 test sequences gave an error rate of 2.9% over a total of 8312 classified data points, where classification error was defined as the difference between inferred segmentation (the estimated switching states) and true segmentation (true switching states) accumulated over all sequences,
Tracking
The SLDS learning framework can improve the tracking of target objects with complex dynamical behavior, responsive to a sequence of measurements. A particular embodiment of the present invention tracks the motion of the human figure, using frames from a video sequence. Tracking systems generally have three identifiable steps: state prediction, measurement processing and posterior update. These steps are exemplified by the well-known Kalman Filter, described in Anderson and Moore, “Optimal Filtering,” Prentice Hall, 1979.
The innovation zt is passed to the posterior update module 502 along with the prediction. They are fused using the Kalman gain matrix to form the posterior estimate {circumflex over (x)}t|t.
The delay element 503 indicates the temporal nature of the tracking process. It models the fact that the posterior estimate for the current frame becomes the input for the filter in the next frame.
We now describe the process of computing the innovation zt which is the task of the measurement processing module 501, in the specific case of figure tracking using template features. A template is a region of pixels, often rectangular in shape. Tracking using template features is described in detail in James M. Rehg and Andrew P. Witkin, “Visual Tracking with Deformation Models,” Proceedings of IEEE Conference on Robotics and Automation, April 1991, Sacramento Calif., pages 844-850. In the figure tracking application, a pixel template is associated with each part of the figure model, describing its image appearance. The pixel template is an example of an “image feature model” which describes the measurements provided by the image.
For example, two templates can be used to describe the right arm, one for the lower and one for the upper arm. These templates, along with those for the left arm and the torso, describe the appearance of the subject's shirt. These templates can be initialized, for example, by capturing pixels from the first frame of the video sequence in which each part is visible. The use of pixel templates in figure tracking is described in detail in Daniel D. Morris and James M. Rehg, “Singularity Analysis for Articulated Object Tracking,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, June 1998, Santa Barbara Calif., pages 289-296.
Given a set of initialized figure templates, the innovation is the difference between the template pixel values and the pixel values in the input video frame which correspond to the predicted template position. We can define the innovation function that gives the pixel difference as a function of the state vector x and pixel index k:
z(x, k)=It(pos(x, k))−T(k) (Eq. 17)
We can then write the innovation at time t as
zt(k)=z({circumflex over (x)}t|t−1, k) (Eq. 18)
Equation 18 defines a vector of pixel differences, zt, formed by subtracting pixels in the template model T from a region of pixels in the input video frame It under the action of the figure state. In order to represent images as vectors, we require that each template pixel in the figure model be assigned a unique index k. This can be easily accomplished by examining the templates in a fixed order and scanning the pixels in each template in raster order. Thus zt(k) represents the innovation for the kth template pixel, whose intensity value is given by T(k). Note that the contents of the template T(k) could be, for example, gray scale pixel values, color pixel values in RGB or YUV, Laplacian filtered pixels, or in general any function applied to a region of pixels. We use z(x) to denote the vector of pixel differences described by Equation (e1). Note that zt does not define an innovation process in the strict sense of the Kalman filter, since the underlying system is nonlinear and non-Gaussian.
The deformation function pos(x, k) in Equation 17 gives the position of the kth template pixel, with respect to the input video frame, as a function of the model state x. Given a state prediction {circumflex over (x)}t|t−1, the transform pos({circumflex over (x)}t|t−1, *) maps the template pixels into the current image, thereby selecting a subset of pixel measurements. In the case of figure tracking, the pos( ) function models the kinematics of the figure, which determine the relative motion of its parts.
Alternatively, the boundary of the figure in the image can be expressed as a collection of image contours whose shapes can be modeled for each part of the figure model. These contours can be measured in an input video frame and the innovation expressed as the distance in the image plane between the positions of predicted and measured contour locations. The use of contour features for tracking is described in detail in Demetri Terzopoulos and Richard Szeliski, “Tracking with Kalman Snakes,” which appears in “Active Vision,” edited by Andrew Blake and Alan Yuille, MIT Press, 1992, pages 3-20. In general, the tracking approach outlined above applies to any set of features which can be computed from a sequence of measurements.
In general, each frame in an image sequence generates a set of “image feature measurements,” such as pixel values, intensity gradients, or edges. Tracking proceeds by fitting an “image feature model,” such as a set of templates or contours, to the set of measurements in each frame.
A primary difficulty in visual tracking is the fact that the image feature measurements are related to the figure state only in a highly nonlinear way. A change in the state of the figure, such as raising an arm, induces a complex change in the pixel values produced in an image. The nonlinear function I(pos(x, *)) models this effect. The standard approach to addressing this nonlinearity is to use the Iterated Extended Kalman Filter (IEKF), which is described in Anderson and Moore, section 8.2. In this approach, the nonlinear measurement model in Equation 17 is linearized around the state prediction {circumflex over (x)}t|t−1.
The linearizing function can be defined as
The first term on the right in Equation 19, ∇It(pos(x, k))′, is the image gradient ∇It, evaluated at the image position of template pixel k. The second term,
is a 2×N kinematic Jacobian matrix, where N is the number of states, or elements, in x. Each column of this Jacobian gives the displacement in the image at pixel position k due to a small change in one of the model states. Mt(x, k) is a 1×N row vector. We denote by Mt(x) the M×N matrix formed by stacking up these rows for each of the M pixels in the template model. The linearized measurement model can then be written:
{overscore (C)}t=Mt({circumflex over (x)}t|t−1) (Eq. 20)
The standard Kalman filter equations can then be applied using the linearized measurement model from Equation 20 and the innovation from Equation 18. In particular, the posterior update equation (analogous to Equation 11 in the linear case) is:
{circumflex over (x)}t|t={circumflex over (x)}t|t−1+Ltzt (Eq. 21)
where Lt is the appropriate Kalman gain matrix formed using {overscore (C)}t.
which result from successive linearizations of the measurement model around the previous posterior estimate
The iterations are initialized by setting
Upon exiting, we set
In cases where the prediction is far from the correct state estimate, these iterations can improve accuracy. We denote as the “IEKF update module” 506 the subsystem of measurement processing 501 and posterior update 502 blocks, as well as delay 504, although in practice, delays 503 and 504 may be the same piece of hardware and/or software.
Clearly the quality of the IEKF solution depends upon the quality of the state prediction. The linearized approximation is only valid within a small range of state values around the current operating point, which is initialized by the prediction. More generally, there are likely to be many background regions in a given video frame that are similar in appearance to the figure. For example, a template that describes the appearance of a figure wearing a dark suit might match quite well to a shadow cast against a wall in the background. The innovation function defined in Equation 17 can be viewed as an objective function to be minimized with respect to x. It is clear that this function will in general have many local minima. The presence of local minima poses problems for any iterative estimation technique such as IEKF. An accurate prediction can reduce the likelihood of becoming trapped in a local minima by bringing the starting point for the iterations closer to the correct answer.
As alluded to previously, the dynamics of a complex moving object such as the human figure cannot be expressed using a simple linear dynamic model. Therefore, it is unlikely that the standard IEKF framework described above would be effective for figure tracking in video. A simple linear prediction of the figure's state would not provide sufficient accuracy in predicting figure motion.
The SLDS framework of the present invention, in contrast, can describe complex dynamics as a switched sequence of linear models. This framework can form an accurate predictor for video tracking. A straightforward implementation of this approach can be extrapolated directly from the approximate Viterbi inference.
In the first tracking embodiment, the Viterbi approach generates a set of S2 hypotheses, corresponding to all possible transitions i, j between linear models from time step t−1 to t. Each of these hypotheses represents a prediction for the continuous state of the figure which selects a corresponding set of pixel measurements. It follows that each hypothesis has a distinct innovation equation:
zt|t−1,i,j=z({circumflex over (x)}t|t−1,i,j) (Eq. 22)
defined by Equation 17. Note that the only difference in comparison to Equation 18 is the use of the SLDS prediction in mapping templates into the image plane.
This embodiment, using SLDS models, is an application of the Viterbi inference algorithm described earlier to the particular measurement models used in visual tracking.
A selector 511 takes these inputs and selects, for each of the S possible switching states at time t, the most likely previous state Ψt−1,j, as defined in Equation 13. This step is identical to the Viterbi inference procedure described earlier, except that the likelihood of the measurement sequence is computed using image features such as templates or contours. More specifically, Equation 7 expresses the transition likelihood Jt|t−1,i,j as the product of the measurement probability and the transition probability from the Markov chain.
In the case of template features, the measurement probability can be written
where {overscore (C)}t|t−1,i,j=Mt({circumflex over (x)}t|t−1,i,j). The difference between Equations 23 and 12 is the use of the linearized measurement model for template features. By changing the measurement probability appropriately, the SLDS framework above can be adopted to any tracking problem.
The output of the selector 511 is a set of S hypotheses corresponding to the most probable state predictions given the measurement data. These are input to the IEKF update block 506, which filters these predictions against the measurements to obtain a set of S posterior state estimates. This block was illustrated in FIG. 15. This step applies the standard equations for the IEKF to features such as templates or contours, as described in Equations 17 through 21. The posterior estimates can be decoded and smoothed analogously to the case of Viterbi inference.
In computing the measurement probabilities, the selector 511 must make S2 comparisons between all of the model pixels and the input image. Depending upon the size of the target and the image, this may represent a large computation. This computation can be reduced by considering only the Markov process probabilities and not the measurement probabilities in computing the best S switching hypotheses. In this case, the best hypotheses are given by
i*t=arg minj{−log Π(i, j)+Jt−1,j} (Eq. 24)
This is a substantial reduction in computation, at the cost of a potential loss of accuracy in picking the best hypotheses.
The selector 511A chooses the most likely transition j→i for each switching state based on the filtered estimates. This is analogous to the Viterbi inference case, which was described in the embodiment of
The switching costs are then given by
Jt,i=maxj{Jt|t,i,jJt−1,j} (Eq. 25A)
where the “posterior transition probability” from state j at time t−1 to state i at time t is written
Jt|t,i,j=P(yt|t,i,j) P(st=ei|st−1=ej) (Eq. 25B)
The selector 511A selects the transition j*→i, where j*=arg maxj{Jt|t,i,j, Jt−1|j}. The state j* is the “optimal posterior switching state,” and its value depends upon the posterior estimate for the continuous state at time t.
The potential advantage of this approach is that all of the S2 predictions are given an opportunity to filter the measurement data. Since the system is nonlinear, it is possible that a hypothesis which has a low probability following prediction could in fact produce a posterior estimate which has a high probability.
The set of S2 predictions used in the embodiments of
When the objective function has many local minima relative to the uncertainty in the state dynamics, it may be advantageous to use additional starting points for search beyond the S2 values provided by the SLDS predictor. This can be accomplished, for example, by drawing additional starting points at random from a “continuous state sampling density.” Additional starting points can increase the chance that one of the IEKF tracks will find the global minimum. We therefore first describe a procedure for generating additional starting points and then describe how they can be incorporated into the previous two embodiments.
A wide range of sampling procedures is available for generating additional starting points for search. In order to apply the IEKF update module 506 at a given starting point, a state prediction must be specified and a particular dynamic model selected. The easiest way to do this is to assume that new starting points are going to be obtained by sampling from the mixture density defined by the set of S2 predictions. This density can be derived as follows:
The first term in the summed product is a Gaussian prior density from the Viterbi inference step. The second term is its likelihood, a scalar probability which we denote αi,j. Rewriting in the form of a mixture of S2 Gaussian kernels, i.e., the predictions, yields:
where αi,j can be further expanded as
Applying Equation 1 to the first term and Equations 5 and 8 to the second term yields
All of the terms in the numerator of Equation 27 are directly available from the Viterbi inference method. The denominator is a constant normalizing factor which can be rewritten to yield
Equations 26 and 28 define a “Viterbi mixture density” from which additional starting points for search can be drawn. The steps for drawing R additional points are as follows:
Find mixture parameters, αi,j, xt|t−1,i,j, and Σt|t−1,i,j from Viterbi prediction For r=1 to R
Select a kernel Ki
Select a state sample xi
End
The rth sample is associated with a specific prediction (ir, jr), making it easy to apply the IEKF.
Given a set of starting points, we can apply the IEKF approach by modifying the IEKF equations to support linearization around arbitrary points. Consider a starting point {overscore (x)}. The nonlinear measurement model for template tracking can be written
zt(xt)=wt
Expanding around {overscore (x)}, discarding high-order terms and rearranging gives
{overscore (y)}t=zt({overscore (x)})−{overscore (C)}t{overscore (x)}=−{overscore (C)}txt+wt
where the auxiliary measurement {overscore (y)}t has been defined to give a measurement model in standard form. It follows that the posterior update equation is
{circumflex over (x)}t|t={circumflex over (x)}t|t−1+Lt(zt({circumflex over (x)})+{overscore (C)}t({circumflex over (x)}t|t−1−{circumflex over (x)})) (Eq. 29)
Note that if {overscore (x)}=xt|t−1, which is the standard operating point for IEKF, then the above reduces to Equation 21. The additional term captures the effect of the new operating point. Equation 29 defines a modification of the IEKF update block to handle arbitrary starting points in computing the posterior update.
Note that the smoothness of the underlying nonlinear measurement model will determine the region in state space over which the linearized model {overscore (C)}t is an accurate approximation. This region must be large enough relative to the distance ∥{circumflex over (x)}t|t−1−{circumflex over (x)}∥. This requirement may not be met by a complex objective function.
An alternative in that case is to apply a standard gradient descent approach instead of IEKF, effectively discounting the role of the prior state prediction. A multiple hypothesis tracking (MHT) approach which implements this procedure is described in Tat-Jen Cham and James M. Rehg, “A Multiple Hypothesis Approach to Figure Tracking,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, June 1999, Ft. Collins, Colo., pages 239-245, incorporated herein by reference in its entirety. Although this article does not describe tracking with an SLDS model or sampling from an SLDS prediction, the MHT procedure provides a means for propagating a set of sampled states given a complex measurement function. This means can be used as an alternative to the IEKF update step.
Along the bottom path in
This combined output forms the input to a final selector 519 which selects one filtered estimate for each of the switching states to make up the final output set. This selection process is identical to the ones described earlier with respect to
It should be clear from the embodiments of
It has been assumed that the selector would be selecting a single most likely state distribution from a set of distributions associated with a particular switching state. Another possibility is to find a single distribution which most closely matches all of the possibilities. This is the Generalized Psuedo Bayesian approximation GPB2, described earlier. In any of the selector blocks for the embodiments discussed above, GPB2 could be used in place of Viterbi approximation as a method for reducing multiple hypotheses for a particular switching state down to a single hypothesis. In cases where no single distribution contains a majority of the probability mass for the set of hypotheses, this approach may be advantageous. The application of the previously described process for GPB2 inference to these embodiments is straightforward.
Synthesis and Interpolation
SLDS was introduced as a “generative” model; trajectories in the state space can be easily generated by simulating an SLDS. Nevertheless, SLDSs are still more commonly employed as a classifier or a filter/predictor than as a generative model. We now formulate a framework for using SLDSs as synthesizers and interpolators.
Consider again the generative model described previously. Provided SLDS parameters have been learned from a corpus of motion trajectories, driving the generative SLDS model with the appropriate state and measurement noise processes and switching model, will yield a state space trajectory consistent with that corpus. In other words, one will draw a sample (trajectory) defined by the probability distribution of the SLDS.
The above procedure will produce a random sequence of samples from the SLDS model. If an average noiseless state trajectory is desired, the synthesis can be run with LDS noise parameters (Q(st), R, Q0(s0)) set to zero and switching states whose duration is equal to average state durations, as determined by the switching state transition matrix Π. For example, this would result in sequences of prototypical walk or jog motions, whereas the random sampling would exhibit deviations from such prototypes. Intermediate levels of randomness can be achieved by scaling the SLDS model noise parameters with factors between 0 and 1.
The model parameters can also be modified to meet new constraints that were, for instance, not present in the data used to learn the SLDS. For example, initial state mean x0, variance Q0 and switching state distribution π can be changed so as to force the SLDS to start in some arbitrary state xa of regime ia, e.g., to start simulation in a “walking” regime ia with figure posture xa. To achieve this, set x0(ia)=xa, Q0(ia)=0, π(ia)=1, and then proceed with the synthesis of this constrained model.
A framework of optimal control can be used to formalize synthesis under constraints. Optimal control of linear dynamic systems is described, for example, in B. Anderson and J. Moore, “Optimal Control: Linear Quadratic Methods,” Prentice Hall, Englewood Cliffs, N.J., 1990. For a LDS, the optimal control framework provides a way to design an optimal input or control ut to the LDS, such that the LDS state xt gets as close as possible to a desired state xtd. The desired states can also be viewed as constraint points. The same formalism can be applied to SLDSs. Namely, the system equation, Equation 1, can be modified as
xt+1=A(s1+1)xt+ut+1+vt+1(st+1) (Eq. 1A)
is minimized, where Wt(x) and Wt(u) are weight matrices. The optimal control is then
For instance, if a SLDS is used to simulate a motion of the human figure, xtd might correspond to a desired figure posture at key frame t, and Wt(x) might designate the key frame, i.e., Wt(x) is large for the key frame, and small otherwise.
In addition to “closeness” constraints, other types of constraints can similarly be added. For example, one can consider a minimum-time constraint where the terminal timer T is to be minimized with the optimal control ût. In that case, the value function to be minimized becomes V(x)←V(x)+T.
Other types of constraints that can be cast in this framework are the inequality or bounding constraints on the state xt or control ut (e.g., xt>xmin, ut<umax). Such constraints could prevent, for example, the limbs of a simulated human figure from assuming physically unrealistic postures, or the control ut from becoming unrealistically large.
If st is known, the solution for the best control ut can be derived using the framework of linear quadratic regulators (LQR) for time-varying LDS. When st is not known, the estimates of st obtained from Viterbi or variational inference can be used and the best control solution can then be found using LQR.
Alternatively, as shown in
Pr(st+1=i|st=j,at+1=k)=Π(a)(i, j, k)
where at represents the control of the switching state at time t, and Π(a)(i, j, k) defines a conditioned switching state transition matrix which depends on the control at 562.
Using the switching control at 562, constraints imposed on the switching states can be satisfied. Such constraints can be formulated similarly to constraints for the continuous control input of
For example, one would like to find at which minimizes the following value function,
Here, ct(st, st−1, at) represents a cost for making the transition from switching state st−1 to st, for a given control at, and γ is a discount or “forgetting” factor. The cost function ct is designed to emphasize, with a low ct, those states that agree with imposed constraints, and to penalize, with a high ct, those states that violate the imposed constraints. Once the optimal control ât has been designed, the modified generative SLDS model can be driven by noise to produce a synthetic trajectory.
As
xt+1=A(st+1)xt+ut+1+v(st+1)
Pr(st+1=i|st=j, at+1=k)=Π(a)(i, j, k)
In this model 570, the mixed control (ut, at) can lead to both a desired switching state, e.g., motion regime, and a desired continuous state, e.g., figure posture. For example, a constraint can be specified that requires the human figure to be in the walking regime id with some specific posture xd at time t. As in the case of the continuous and switching state optimal controls, additional constraints such as input bounding and minimum time can also be specified for the mixed state control. To find the optimal control (ût, ât), a value function can be used that includes the costs of the switching and the continuous state controls, e.g., V(x)+V(s). Again, once the optimal control (ût, ât) is designed, the modified generative SLDS model can be used to produce a synthetic trajectory.
The use of optimal controls by the present invention to generate motion by sampling from SLDS models in the presence of constraints has broad applications in computer animation and computer graphics. The task of generating realistic or compelling motions for synthetic characters has long been recognized as an extremely challenging problem. One classical approach to this problem is based on the notion of spacetime constraints which was first introduced by Andrew Witkin and Michael Kass, “Spacetime Constraints,” Computer Graphics, Volume 22, Number 4, August 1988 pages 159-168. In this approach, an optimal control problem is formulated over an analytic dynamical model derived from Newtonian physics. For example, in order to make a lamp hop in a realistic way, the animator would derive the physical equations which govern the lamp's motion. These equations involve the mass distribution of the lamp and forces such as gravity that are acting upon it.
Unfortunately, this method of animation has proved to be extremely difficult to use in practice. There are two main problems. First, it is extremely difficult to specify all of the equations and parameters that are necessary to produce a desired motion. In the case of the human figure, for example, specifying all of the model parameters required for realistic motion is a daunting task. The second problem is that the resulting equations of motion are highly complex and nonlinear, and usually involve an enormous number of state variables. For example, the jumping Luxo lamp described in the Witkin and Kass paper involved 223 constraints and 394 state variables. The numerical methods which must be used to solve control problems in such a large complex state space are also difficult to work with. Their implementation is complex and their convergence to a correct solution can be problematic.
In contrast, our approach of optimal control of learned SLDS models has the potential to overcome both of these drawbacks. By using a learned switching model, desired attributes such as realism can be obtained directly from motion data. Thus, there is no need to specify the physics of human motion explicitly in order to obtain useful motion synthesis. By incorporating additional constraints through the mechanism of optimal control, we enable an animator to achieve specific artistic goals by controlling the synthesized motion. One potential advantage of our optimal control framework over the classical spacetime approach is the linearity of the SLDS model once a switching sequence has been specified. This makes it possible to use optimal control techniques such as the linear quadratic regulator which are extremely stable and well-understood. Implementations of LQR can be found in many standard software packages such as Matlab. This stands in contrast to the sequential quadratic programming methods required by the classical spacetime approach.
The spacetime approach can be extended to include the direct use of motion capture data. This is described in Michael Gleicher, “Retargetting Motion to New Characters,” Proceedings of SIGGRAPH 98, in Computer Graphics Proceedings, Annual Conference series, 1998, pages 33-42. In this method, a single sequence of human motion data is modified to achieve a specific animation objective by filtering it with a biomechanical model of human motion and adjusting the model parameters. With this method, a motion sequence of a person walking can be adapted to a figure model whose proportions are different from that of the subject. Motions can also be modified to satisfy various equality or inequality constraints.
The use of motion capture data within the spacetime framework makes it possible to achieve realistic motion without the complete specification of biomechanical model parameters. However this approach still suffers from the complexity of the underlying numerical methods. Another disadvantage of this approach in contrast to the current invention is that there is no obvious way to generate multiple examples of the same type of motion or to generate new motions by combining several examples of motion data.
In contrast, sampling repeatedly from our SLDS framework can produce motions which differ in their small details but are qualitatively consistent. This type of randomness is necessary in order to avoid awkward repetitions in an animation application. Furthermore, the learning approach makes it possible to generalize from a set of motions to new motions that have not been seen before. In contrast, the approach of retargeting is based on modifying a single instance of motion data.
Another approach to synthesizing animations from learned models is described in Matthew Brand, “Voice Puppetry,” SIGGRAPH 99 Conference Proceedings, Annual Conference Series 1999, pages 21-28 and in Matthew Brand and Aaron Hertzmann, “Style Machines,” SIGGRAPH 2000 Conference Proceedings, Annual Conference Series 2000, pages 183-192. This method uses sampling from a Hidden Markov Model to synthesize facial and figure animations learned from training data. Unlike our SLDS framework, the HMM representation is limited to using piecewise constant functions to model the feature data during learning. This can require a large number of discrete states in order to capture subtle effects when working with complex state space data.
In contrast, our framework employs a set of LDS models to describe feature data, resulting in models with much more expressive power. Furthermore, the prior art does not describe any mechanism for imposing constraints on the samples from the models. This may make it difficult to use this approach in achieving specific animation objectives.
To test the power of the learned SLDS synthesis/interpolation framework, we examined its use in synthesizing realistic-looking motion sequences and interpolating motion between missing frames. In one set of experiments, the learned walk/jog SLDS was used to generate a “synthetic walk” based on initial conditions learned by the SLDS model.
Another realistic situation may call for filling in a small number of missing frames from a large motion sequence. SLDS can then be utilized as an interpolation function. In another set of experiments, we employed the learned walk/jog model to interpolate a walk motion over two sequences with missing frames. Missing-frame constraints were included in the interpolation framework by setting the measurement variances corresponding to those frames to infinity. The visual quality of the interpolation and the motion synthesized from it was high. As expected, the sparseness of the measurement set had definite bearing on this quality.
Use of the Invention
Our invention makes possible a number of core tasks related to the analysis and synthesis of the human figure motion:
Track figure motion in an image sequence using learned dynamic models.
Classify different types of human motion.
Synthesize motion using stochastic models that correspond to different types of motion.
Interpolate missing motion data from sparsely observed image sequences.
We anticipate that our invention could impact the following application areas:
Surveillance: Use of accurate dynamic models could lead to improved tracking in noisy video footage. The ability to interpolate missing data could be useful in situations where frame rates are low, as in Web or other network applications. The ability to classify motion into categories can benefit from the SLDS approach. Two forms of classification are possible. First, specific actions such as “opening a door” or “dropping a package” can be modeled using our approach. Second, it may be possible to recognize specific individuals based on the observed dynamics of their motion in image sequences. This could be used, for example, to recognize criminal suspects using surveillance cameras in public places such as airports or train stations.
User-interfaces: Interfaces based on vision sensing could benefit from improved tracking and better classification performance due to the SLDS approach.
Motion capture: Motion capture in unstructured environments can be enabled through better tracking techniques. In addition to the capture of live motion without the use of special clothing, it is also possible to capture motion from archival sources such as old movies.
Motion synthesis: The generation of human motion for computer graphics animation can be enabled through the use of a learned, generative stochastic model. By learning models from sample motions, it is possible to capture the natural dynamics implicit in real human motion without a laborious manual modeling process. Because the resulting models are stochastic, sampling from the models produces motion with a pleasing degree of randomness.
Video editing: Tracking algorithms based on powerful dynamic models can simplify the task of segmenting video sequences.
Video compression/decompression: The ability to interpolate a video sequence based on a sparse set of samples could provide a new approach to coding and decoding video sequences containing human or other motion. In practice, human motion is common in video sequences. By transmitting key frames detected using SLSDS classification at a low sampling rate and interpolating, using SLDS interpolation, the missing frames from the transmitted model parameters, a substantial savings in bit-rate may be achievable.
It will be apparent to those of ordinary skill in the art that methods involved in the present system for classifying sequences may be embodied in a computer program product that includes a computer usable medium. For example, such a computer usable medium can include a readable memory device, such as a hard drive device, a CD-ROM, a DVD-ROM, or a computer diskette, having computer readable program code segments stored thereon. The computer readable medium can also include a communications or transmission medium, such as a bus or a communications link, either optical, wired, or wireless, having program code segments carried thereon as digital or analog data signals.
While this invention has been particularly shown and described with references to preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the scope of the invention encompassed by the appended claims.
This application is a continuation of U.S. application Ser. No. 09/654,300, filed Sep. 1, 2000, now U.S. Pat. No. 6,694,044, which claims the benefit of U.S. Provisional Application No. 60/154,384, filed on Sep. 16, 1999. The entire teachings of the above applications are incorporated herein by reference.
Number | Name | Date | Kind |
---|---|---|---|
5325098 | Blair et al. | Jun 1994 | A |
5923712 | Leyendecker et al. | Jul 1999 | A |
6064703 | Cusani et al. | May 2000 | A |
6226409 | Cham et al. | May 2001 | B1 |
6243037 | Pulford et al. | Jun 2001 | B1 |
6256418 | Rehg et al. | Jul 2001 | B1 |
6314204 | Cham et al. | Nov 2001 | B1 |
6393046 | Kohli et al. | May 2002 | B1 |
6396878 | Piirainen | May 2002 | B1 |
6480876 | Rehg et al. | Nov 2002 | B2 |
Number | Date | Country | |
---|---|---|---|
20040042639 A1 | Mar 2004 | US |
Number | Date | Country | |
---|---|---|---|
60154384 | Sep 1999 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 09654300 | Sep 2000 | US |
Child | 10663938 | US |