Stochastic Human Segmentation from a Static Camera ￡

合集下载

Conditional Random Fields_ Probabilistic Models for Segmenting and Labeling Sequence Data

Conditional Random Fields:Probabilistic Modelsfor Segmenting and Labeling Sequence DataJohn Lafferty LAFFERTY@ Andrew McCallum MCCALLUM@ Fernando Pereira FPEREIRA@ WhizBang!Labs–Research,4616Henry Street,Pittsburgh,PA15213USASchool of Computer Science,Carnegie Mellon University,Pittsburgh,PA15213USADepartment of Computer and Information Science,University of Pennsylvania,Philadelphia,PA19104USAAbstractWe present,a frame-work for building probabilistic models to seg-ment and label sequence data.Conditional ran-domﬁelds offer several advantages over hid-den Markov models and stochastic grammarsfor such tasks,including the ability to relaxstrong independence assumptions made in thosemodels.Conditional randomﬁelds also avoida fundamental limitation of maximum entropyMarkov models(MEMMs)and other discrimi-native Markov models based on directed graph-ical models,which can be biased towards stateswith few successor states.We present iterativeparameter estimation algorithms for conditionalrandomﬁelds and compare the performance ofthe resulting models to HMMs and MEMMs onsynthetic and natural-language data.1.IntroductionThe need to segment and label sequences arises in many different problems in several scientiﬁcﬁelds.Hidden Markov models(HMMs)and stochastic grammars are well understood and widely used probabilistic models for such problems.In computational biology,HMMs and stochas-tic grammars have been successfully used to align bio-logical sequences,ﬁnd sequences homologous to a known evolutionary family,and analyze RNA secondary structure (Durbin et al.,1998).In computational linguistics and computer science,HMMs and stochastic grammars have been applied to a wide variety of problems in text and speech processing,including topic segmentation,part-of-speech(POS)tagging,information extraction,and syntac-tic disambiguation(Manning&Sch¨u tze,1999).HMMs and stochastic grammars are generative models,as-signing a joint probability to paired observation and label sequences;the parameters are typically trained to maxi-mize the joint likelihood of training examples.To deﬁne a joint probability over observation and label sequences, a generative model needs to enumerate all possible ob-servation sequences,typically requiring a representation in which observations are task-appropriate atomic entities, such as words or nucleotides.In particular,it is not practi-cal to represent multiple interacting features or long-range dependencies of the observations,since the inference prob-lem for such models is intractable.This difﬁculty is one of the main motivations for looking at conditional models as an alternative.A conditional model speciﬁes the probabilities of possible label sequences given an observation sequence.Therefore,it does not expend modeling effort on the observations,which at test time areﬁxed anyway.Furthermore,the conditional probabil-ity of the label sequence can depend on arbitrary,non-independent features of the observation sequence without forcing the model to account for the distribution of those dependencies.The chosen features may represent attributes at different levels of granularity of the same observations (for example,words and characters in English text),or aggregate properties of the observation sequence(for in-stance,text layout).The probability of a transition between labels may depend not only on the current observation, but also on past and future observations,if available.In contrast,generative models must make very strict indepen-dence assumptions on the observations,for instance condi-tional independence given the labels,to achieve tractability. Maximum entropy Markov models(MEMMs)are condi-tional probabilistic sequence models that attain all of the above advantages(McCallum et al.,2000).In MEMMs, each source state1has a exponential model that takes the observation features as input,and outputs a distribution over possible next states.These exponential models are trained by an appropriate iterative scaling method in the 1Output labels are associated with states;it is possible for sev-eral states to have the same label,but for simplicity in the rest of this paper we assume a one-to-one correspondence.maximum entropy framework.Previously published exper-imental results show MEMMs increasing recall and dou-bling precision relative to HMMs in a FAQ segmentation task.MEMMs and other non-generativeﬁnite-state models based on next-state classiﬁers,such as discriminative Markov models(Bottou,1991),share a weakness we callhere the:the transitions leaving a given state compete only against each other,rather than againstall other transitions in the model.In probabilistic terms, transition scores are the conditional probabilities of pos-sible next states given the current state and the observa-tion sequence.This per-state normalization of transition scores implies a“conservation of score mass”(Bottou, 1991)whereby all the mass that arrives at a state must be distributed among the possible successor states.An obser-vation can affect which destination states get the mass,but not how much total mass to pass on.This causes a bias to-ward states with fewer outgoing transitions.In the extreme case,a state with a single outgoing transition effectively ignores the observation.In those cases,unlike in HMMs, Viterbi decoding cannot downgrade a branch based on ob-servations after the branch point,and models with state-transition structures that have sparsely connected chains of states are not properly handled.The Markovian assump-tions in MEMMs and similar state-conditional models in-sulate decisions at one state from future decisions in a way that does not match the actual dependencies between con-secutive states.This paper introduces(CRFs),a sequence modeling framework that has all the advantages of MEMMs but also solves the label bias problem in a principled way.The critical difference between CRFs and MEMMs is that a MEMM uses per-state exponential mod-els for the conditional probabilities of next states given the current state,while a CRF has a single exponential model for the joint probability of the entire sequence of labels given the observation sequence.Therefore,the weights of different features at different states can be traded off against each other.We can also think of a CRF as aﬁnite state model with un-normalized transition probabilities.However,unlike some other weightedﬁnite-state approaches(LeCun et al.,1998), CRFs assign a well-deﬁned probability distribution over possible labelings,trained by maximum likelihood or MAP estimation.Furthermore,the loss function is convex,2guar-anteeing convergence to the global optimum.CRFs also generalize easily to analogues of stochastic context-free grammars that would be useful in such problems as RNA secondary structure prediction and natural language pro-cessing.2In the case of fully observable states,as we are discussing here;if several states have the same label,the usual local maxima of Baum-Welch arise.bel bias example,after(Bottou,1991).For concise-ness,we place observation-label pairs on transitions rather than states;the symbol‘’represents the null output label.We present the model,describe two training procedures and sketch a proof of convergence.We also give experimental results on synthetic data showing that CRFs solve the clas-sical version of the label bias problem,and,more signiﬁ-cantly,that CRFs perform better than HMMs and MEMMs when the true data distribution has higher-order dependen-cies than the model,as is often the case in practice.Finally, we conﬁrm these results as well as the claimed advantages of conditional models by evaluating HMMs,MEMMs and CRFs with identical state structure on a part-of-speech tag-ging task.2.The Label Bias ProblemClassical probabilistic automata(Paz,1971),discrimina-tive Markov models(Bottou,1991),maximum entropy taggers(Ratnaparkhi,1996),and MEMMs,as well as non-probabilistic sequence tagging and segmentation mod-els with independently trained next-state classiﬁers(Pun-yakanok&Roth,2001)are all potential victims of the label bias problem.For example,Figure1represents a simpleﬁnite-state model designed to distinguish between the two words and.Suppose that the observation sequence is. In theﬁrst time step,matches both transitions from the start state,so the probability mass gets distributed roughly equally among those two transitions.Next we observe. Both states1and4have only one outgoing transition.State 1has seen this observation often in training,state4has al-most never seen this observation;but like state1,state4 has no choice but to pass all its mass to its single outgoing transition,since it is not generating the observation,only conditioning on it.Thus,states with a single outgoing tran-sition effectively ignore their observations.More generally, states with low-entropy next state distributions will take lit-tle notice of observations.Returning to the example,the top path and the bottom path will be about equally likely, independently of the observation sequence.If one of the two words is slightly more common in the training set,the transitions out of the start state will slightly prefer its cor-responding transition,and that word’s state sequence will always win.This behavior is demonstrated experimentally in Section5.L´e on Bottou(1991)discussed two solutions for the label bias problem.One is to change the state-transition struc-ture of the model.In the above example we could collapse states1and4,and delay the branching until we get a dis-criminating observation.This operation is a special case of determinization(Mohri,1997),but determinization of weightedﬁnite-state machines is not always possible,and even when possible,it may lead to combinatorial explo-sion.The other solution mentioned is to start with a fully-connected model and let the training procedureﬁgure out a good structure.But that would preclude the use of prior structural knowledge that has proven so valuable in infor-mation extraction tasks(Freitag&McCallum,2000). Proper solutions require models that account for whole state sequences at once by letting some transitions“vote”more strongly than others depending on the corresponding observations.This implies that score mass will not be con-served,but instead individual transitions can“amplify”or “dampen”the mass they receive.In the above example,the transitions from the start state would have a very weak ef-fect on path score,while the transitions from states1and4 would have much stronger effects,amplifying or damping depending on the actual observation,and a proportionally higher contribution to the selection of the Viterbi path.3In the related work section we discuss other heuristic model classes that account for state sequences globally rather than locally.To the best of our knowledge,CRFs are the only model class that does this in a purely probabilistic setting, with guaranteed global maximum likelihood convergence.3.Conditional Random FieldsIn what follows,is a random variable over data se-quences to be labeled,and is a random variable over corresponding label sequences.All components of are assumed to range over aﬁnite label alphabet.For ex-ample,might range over natural language sentences and range over part-of-speech taggings of those sentences, with the set of possible part-of-speech tags.The ran-dom variables and are jointly distributed,but in a dis-criminative framework we construct a conditional model from paired observation and label sequences,and do not explicitly model the marginal..Thus,a CRF is a randomﬁeld globally conditioned on the observation.Throughout the paper we tacitly assume that the graph isﬁxed.In the simplest and most impor-3Weighted determinization and minimization techniques shift transition weights while preserving overall path weight(Mohri, 2000);their connection to this discussion deserves further study.tant example for modeling sequences,is a simple chain or line:.may also have a natural graph structure;yet in gen-eral it is not necessary to assume that and have the same graphical structure,or even that has any graph-ical structure at all.However,in this paper we will be most concerned with sequencesand.If the graph of is a tree(of which a chain is the simplest example),its cliques are the edges and ver-tices.Therefore,by the fundamental theorem of random ﬁelds(Hammersley&Clifford,1971),the joint distribu-tion over the label sequence given has the form(1),where is a data sequence,a label sequence,and is the set of components of associated with the vertices in subgraph.We assume that the and are given andﬁxed. For example,a Boolean vertex feature might be true if the word is upper case and the tag is“proper noun.”The parameter estimation problem is to determine the pa-rameters from training datawith empirical distribution. In Section4we describe an iterative scaling algorithm that maximizes the log-likelihood objective function:.As a particular case,we can construct an HMM-like CRF by deﬁning one feature for each state pair,and one feature for each state-observation pair:. The corresponding parameters and play a simi-lar role to the(logarithms of the)usual HMM parameters and.Boltzmann chain models(Saul&Jor-dan,1996;MacKay,1996)have a similar form but use a single normalization constant to yield a joint distribution, whereas CRFs use the observation-dependent normaliza-tion for conditional distributions.Although it encompasses HMM-like models,the class of conditional randomﬁelds is much more expressive,be-cause it allows arbitrary dependencies on the observationFigure2.Graphical structures of simple HMMs(left),MEMMs(center),and the chain-structured case of CRFs(right)for sequences. An open circle indicates that the variable is not generated by the model.sequence.In addition,the features do not need to specify completely a state or observation,so one might expect that the model can be estimated from less training data.Another attractive property is the convexity of the loss function;in-deed,CRFs share all of the convexity properties of general maximum entropy models.For the remainder of the paper we assume that the depen-dencies of,conditioned on,form a chain.To sim-plify some expressions,we add special start and stop states and.Thus,we will be using the graphical structure shown in Figure2.For a chain struc-ture,the conditional probability of a label sequence can be expressed concisely in matrix form,which will be useful in describing the parameter estimation and inference al-gorithms in Section4.Suppose that is a CRF given by(1).For each position in the observation se-quence,we deﬁne the matrix random variableby,where is the edge with labels and is the vertex with label.In contrast to generative models,con-ditional models like CRFs do not need to enumerate over all possible observation sequences,and therefore these matrices can be computed directly as needed from a given training or test observation sequence and the parameter vector.Then the normalization(partition function)is the entry of the product of these matrices:. Using this notation,the conditional probability of a label sequence is written as, where and.4.Parameter Estimation for CRFsWe now describe two iterative scaling algorithms toﬁnd the parameter vector that maximizes the log-likelihood of the training data.Both algorithms are based on the im-proved iterative scaling(IIS)algorithm of Della Pietra et al. (1997);the proof technique based on auxiliary functions can be extended to show convergence of the algorithms for CRFs.Iterative scaling algorithms update the weights asand for appropriately chosen and.In particular,the IIS update for an edge feature is the solution ofdef.where is thedef. The equations for vertex feature updates have similar form.However,efﬁciently computing the exponential sums on the right-hand sides of these equations is problematic,be-cause is a global property of,and dynamic programming will sum over sequences with potentially varying.To deal with this,theﬁrst algorithm,Algorithm S,uses a“slack feature.”The second,Algorithm T,keepstrack of partial totals.For Algorithm S,we deﬁne the bydef,where is a constant chosen so that for all and all observation vectors in the training set,thus making.Feature is“global,”that is,it does not correspond to any particular edge or vertex.For each index we now deﬁne thewith base caseifotherwiseand recurrence.Similarly,the are deﬁned byifotherwiseand.With these deﬁnitions,the update equations are,where.The factors involving the forward and backward vectors in the above equations have the same meaning as for standard hidden Markov models.For example,is the marginal probability of label given that the observation sequence is.This algorithm is closely related to the algorithm of Darroch and Ratcliff(1972),and MART algorithms used in image reconstruction.The constant in Algorithm S can be quite large,since in practice it is proportional to the length of the longest train-ing observation sequence.As a result,the algorithm may converge slowly,taking very small steps toward the maxi-mum in each iteration.If the length of the observations and the number of active features varies greatly,a faster-converging algorithm can be obtained by keeping track of feature totals for each observation sequence separately. Let def.Algorithm T accumulates feature expectations into counters indexed by.More speciﬁcally,we use the forward-backward recurrences just introduced to compute the expectations of feature and of feature given that.Then our param-eter updates are and,whereand are the unique positive roots to the following polynomial equationsmax max,(2)which can be easily computed by Newton’s method.A single iteration of Algorithm S and Algorithm T has roughly the same time and space complexity as the well known Baum-Welch algorithm for HMMs.To prove con-vergence of our algorithms,we can derive an auxiliary function to bound the change in likelihood from below;this method is developed in detail by Della Pietra et al.(1997). The full proof is somewhat detailed;however,here we give an idea of how to derive the auxiliary function.To simplify notation,we assume only edge features with parameters .Given two parameter settings and,we bound from below the change in the objective function with anas followsdefwhere the inequalities follow from the convexity of and.Differentiating with respect to and setting the result to zero yields equation(2).5.ExperimentsWeﬁrst discuss two sets of experiments with synthetic data that highlight the differences between CRFs and MEMMs. Theﬁrst experiments are a direct veriﬁcation of the label bias problem discussed in Section2.In the second set of experiments,we generate synthetic data using randomly chosen hidden Markov models,each of which is a mix-ture of aﬁrst-order and second-order peting models are then trained and compared on test data.As the data becomes more second-order,the test er-ror rates of the trained models increase.This experiment corresponds to the common modeling practice of approxi-mating complex local and long-range dependencies,as oc-cur in natural data,by small-order Markov models.OurFigure3.Plots of error rates for HMMs,CRFs,and MEMMs on randomly generated synthetic data sets,as described in Section5.2. As the data becomes“more second order,”the error rates of the test models increase.As shown in the left plot,the CRF typically signiﬁcantly outperforms the MEMM.The center plot shows that the HMM outperforms the MEMM.In the right plot,each open square represents a data set with,and a solid circle indicates a data set with.The plot shows that when the data is mostly second order(),the discriminatively trained CRF typically outperforms the HMM.These experiments are not designed to demonstrate the advantages of the additional representational power of CRFs and MEMMs relative to HMMs.results clearly indicate that even when the models are pa-rameterized in exactly the same way,CRFs are more ro-bust to inaccurate modeling assumptions than MEMMs or HMMs,and resolve the label bias problem,which affects the performance of MEMMs.To avoid confusion of dif-ferent effects,the MEMMs and CRFs in these experiments use overlapping features of the observations.Fi-nally,in a set of POS tagging experiments,we conﬁrm the advantage of CRFs over MEMMs.We also show that the addition of overlapping features to CRFs and MEMMs al-lows them to perform much better than HMMs,as already shown for MEMMs by McCallum et al.(2000).5.1Modeling label biasWe generate data from a simple HMM which encodes a noisy version of theﬁnite-state network in Figure1.Each state emits its designated symbol with probabilityand any of the other symbols with probability.We train both an MEMM and a CRF with the same topologies on the data generated by the HMM.The observation fea-tures are simply the identity of the observation symbols. In a typical run using training and test samples, trained to convergence of the iterative scaling algorithm, the CRF error is while the MEMM error is, showing that the MEMM fails to discriminate between the two branches.5.2Modeling mixed-order sourcesFor these results,we useﬁve labels,(),and26 observation values,();however,the results were qualitatively the same over a range of sizes for and .We generate data from a mixed-order HMM with state transition probabilities given byand,simi-larly,emission probabilities given by.Thus,for we have a standardﬁrst-order HMM.In order to limit the size of the Bayes error rate for the resulting models,the con-ditional probability tables are constrained to be sparse. In particular,can have at most two nonzero en-tries,for each,and can have at most three nonzero entries for each.For each randomly gener-ated model,a sample of1,000sequences of length25is generated for training and testing.On each randomly generated training set,a CRF is trained using Algorithm S.(Note that since the length of the se-quences and number of active features is constant,Algo-rithms S and T are identical.)The algorithm is fairly slow to converge,typically taking approximately500iterations for the model to stabilize.On the500MHz Pentium PC used in our experiments,each iteration takes approximately 0.2seconds.On the same data an MEMM is trained using iterative scaling,which does not require forward-backward calculations,and is thus more efﬁcient.The MEMM train-ing converges more quickly,stabilizing after approximately 100iterations.For each model,the Viterbi algorithm is used to label a test set;the experimental results do not sig-niﬁcantly change when using forward-backward decoding to minimize the per-symbol error rate.The results of several runs are presented in Figure3.Each plot compares two classes of models,with each point indi-cating the error rate for a single test set.As increases,the error rates generally increase,as theﬁrst-order models fail toﬁt the second-order data.Theﬁgure compares models parameterized as,,and;results for models parameterized as,,and are qualitatively the same.As shown in theﬁrst graph,the CRF generally out-performs the MEMM,often by a wide margin of10%–20% relative error.(The points for very small error rate,with ,where the MEMM does better than the CRF, are suspected to be the result of an insufﬁcient number of training iterations for the CRF.)HMM 5.69%45.99%MEMM 6.37%54.61%CRF 5.55%48.05%MEMM 4.81%26.99%CRF 4.27%23.76%Using spelling featuresFigure4.Per-word error rates for POS tagging on the Penn tree-bank,usingﬁrst-order models trained on50%of the1.1million word corpus.The oov rate is5.45%.5.3POS tagging experimentsTo conﬁrm our synthetic data results,we also compared HMMs,MEMMs and CRFs on Penn treebank POS tag-ging,where each word in a given input sentence must be labeled with one of45syntactic tags.We carried out two sets of experiments with this natural language data.First,we trainedﬁrst-order HMM,MEMM, and CRF models as in the synthetic data experiments,in-troducing parameters for each tag-word pair andfor each tag-tag pair in the training set.The results are con-sistent with what is observed on synthetic data:the HMM outperforms the MEMM,as a consequence of the label bias problem,while the CRF outperforms the HMM.The er-ror rates for training runs using a50%-50%train-test split are shown in Figure5.3;the results are qualitatively sim-ilar for other splits of the data.The error rates on out-of-vocabulary(oov)words,which are not observed in the training set,are reported separately.In the second set of experiments,we take advantage of the power of conditional models by adding a small set of or-thographic features:whether a spelling begins with a num-ber or upper case letter,whether it contains a hyphen,and whether it ends in one of the following sufﬁxes:.Here weﬁnd,as expected,that both the MEMM and the CRF beneﬁt signif-icantly from the use of these features,with the overall error rate reduced by around25%,and the out-of-vocabulary er-ror rate reduced by around50%.One usually starts training from the all zero parameter vec-tor,corresponding to the uniform distribution.However, for these datasets,CRF training with that initialization is much slower than MEMM training.Fortunately,we can use the optimal MEMM parameter vector as a starting point for training the corresponding CRF.In Figure5.3, MEMM was trained to convergence in around100iter-ations.Its parameters were then used to initialize the train-ing of CRF,which converged in1,000iterations.In con-trast,training of the same CRF from the uniform distribu-tion had not converged even after2,000iterations.6.Further Aspects of CRFsMany further aspects of CRFs are attractive for applica-tions and deserve further study.In this section we brieﬂy mention just two.Conditional randomﬁelds can be trained using the expo-nential loss objective function used by the AdaBoost algo-rithm(Freund&Schapire,1997).Typically,boosting is applied to classiﬁcation problems with a small,ﬁxed num-ber of classes;applications of boosting to sequence labeling have treated each label as a separate classiﬁcation problem (Abney et al.,1999).However,it is possible to apply the parallel update algorithm of Collins et al.(2000)to op-timize the per-sequence exponential loss.This requires a forward-backward algorithm to compute efﬁciently certain feature expectations,along the lines of Algorithm T,ex-cept that each feature requires a separate set of forward and backward accumulators.Another attractive aspect of CRFs is that one can imple-ment efﬁcient feature selection and feature induction al-gorithms for them.That is,rather than specifying in ad-vance which features of to use,we could start from feature-generating rules and evaluate the beneﬁt of gener-ated features automatically on data.In particular,the fea-ture induction algorithms presented in Della Pietra et al. (1997)can be adapted toﬁt the dynamic programming techniques of conditional randomﬁelds.7.Related Work and ConclusionsAs far as we know,the present work is theﬁrst to combine the beneﬁts of conditional models with the global normal-ization of randomﬁeld models.Other applications of expo-nential models in sequence modeling have either attempted to build generative models(Rosenfeld,1997),which in-volve a hard normalization problem,or adopted local con-ditional models(Berger et al.,1996;Ratnaparkhi,1996; McCallum et al.,2000)that may suffer from label bias. Non-probabilistic local decision models have also been widely used in segmentation and tagging(Brill,1995; Roth,1998;Abney et al.,1999).Because of the computa-tional complexity of global training,these models are only trained to minimize the error of individual label decisions assuming that neighboring labels are correctly -bel bias would be expected to be a problem here too.An alternative approach to discriminative modeling of se-quence labeling is to use a permissive generative model, which can only model local dependencies,to produce a list of candidates,and then use a more global discrimina-tive model to rerank those candidates.This approach is standard in large-vocabulary speech recognition(Schwartz &Austin,1993),and has also been proposed for parsing (Collins,2000).However,these methods fail when the cor-rect output is pruned away in theﬁrst pass.。

初三英语哲学思考问题单选题40题

初三英语哲学思考问题单选题40题1. When we think about the nature of reality, which of the following statements is correct?A. Reality is only what we can see.B. Reality is determined by our thoughts.C. Reality is independent of human perception.D. Reality changes based on our feelings.答案：C。

本题主要考查对现实本质的哲学理解。

选项A 过于局限，现实不仅仅是我们能看到的。

选项B 是主观唯心主义观点，不符合客观事实。

选项C 符合唯物主义观点，现实是独立于人类感知而存在的。

选项D 现实不会仅仅因为我们的感受而改变。

2. What is the essence of philosophy according to the basic concepts?A. The study of history.B. The exploration of science.C. The reflection on fundamental questions of life and existence.D. The analysis of language.答案：C。

哲学的本质是对生命和存在的基本问题进行反思。

选项 A 历史研究并非哲学的本质。

选项 B 科学探索也不是哲学的本质核心。

选项D 语言分析只是哲学的一个方面，而非本质。

3. In the philosophical view, which one is true about truth?A. Truth is relative and changes over time.B. Truth is absolute and never changes.C. Truth depends on personal belief.D. Truth is something that cannot be known.答案：A。

中国科学英文版模板

中国科学英文版模板1.Identification of Wiener systems with nonlinearity being piece wise-linear function HUANG YiQing，CHEN HanFu，FANG HaiTao2.A novel algorithm for explicit optimal multi-degree reduction of triangular surfaces HU QianQian，WANG GuoJin3.New approach to the automatic segmentation of coronary arte ry in X-ray angiograms ZHOU ShouJun，YANG Jun，CHEN WuFan，WANG YongTian4.Novel Ω-protocols for NP DENG Yi，LIN DongDai5.Non-coherent space-time code based on full diversity space-ti me block coding GUO YongLiang，ZHU ShiHua6.Recursive algorithm and accurate computation of dyadic Green 's functions for stratified uniaxial anisotropic media WEI BaoJun，ZH ANG GengJi，LIU QingHuo7.A blind separation method of overlapped multi-components b ased on time varying AR model CAI QuanWei，WEI Ping，XIAO Xian Ci8.Joint multiple parameters estimation for coherent chirp signals using vector sensor array WEN Zhong，LI LiPing，CHEN TianQi，ZH ANG XiXiang9.Vision implants: An electrical device will bring light to the blind NIU JinHai，LIU YiFei，REN QiuShi，ZHOU Yang，ZHOU Ye，NIU S huaibining search space partition and search Space partition and ab straction for LTL model checking PU Fei，ZHANG WenHui2.Dynamic replication of Web contents Amjad Mahmood3.On global controllability of affine nonlinear systems with a tria ngular-like structure SUN YiMin，MEI ShengWei，LU Qiang4.A fuzzy model of predicting RNA secondary structure SONG D anDan，DENG ZhiDong5.Randomization of classical inference patterns and its applicatio n WANG GuoJun，HUI XiaoJing6.Pulse shaping method to compensate for antenna distortion in ultra-wideband communications WU XuanLi，SHA XueJun，ZHANG NaiTong7.Study on modulation techniques free of orthogonality restricti on CAO QiSheng，LIANG DeQun8.Joint-state differential detection algorithm and its application in UWB wireless communication systems ZHANG Peng，BI GuangGuo，CAO XiuYing9.Accurate and robust estimation of phase error and its uncertai nty of 50 GHz bandwidth sampling circuit ZHANG Zhe，LIN MaoLiu，XU QingHua，TAN JiuBin10.Solving SAT problem by heuristic polarity decision-making al gorithm JING MingE，ZHOU Dian，TANG PuShan，ZHOU XiaoFang，ZHANG Hua1.A novel formal approach to program slicing ZHANG YingZhou2.On Hamiltonian realization of time-varying nonlinear systems WANG YuZhen，Ge S. S.，CHENG DaiZhan3.Primary exploration of nonlinear information fusion control the ory WANG ZhiSheng，WANG DaoBo，ZHEN ZiYang4.Center-configur ation selection technique for the reconfigurable modular robot LIU J inGuo，WANG YueChao，LI Bin，MA ShuGen，TAN DaLong5.Stabilization of switched linear systems with bounded disturba nces and unobservable switchings LIU Feng6.Solution to the Generalized Champagne Problem on simultane ous stabilization of linear systems GUAN Qiang，WANG Long，XIA B iCan，YANG Lu，YU WenSheng，ZENG ZhenBing7.Supporting service differentiation with enhancements of the IE EE 802.11 MAC protocol: Models and analysis LI Bo，LI JianDong，R oberto Battiti8.Differential space-time block-diagonal codes LUO ZhenDong，L IU YuanAn，GAO JinChun9.Cross-layer optimization in ultra wideband networks WU Qi，BI JingPing，GUO ZiHua，XIONG YongQiang，ZHANG Qian，LI ZhongC heng10.Searching-and-averaging method of underdetermined blind s peech signal separation in time domain XIAO Ming，XIE ShengLi，F U YuLi11.New theoretical framework for OFDM/CDMA systems with pe ak-limited nonlinearities WANG Jian，ZHANG Lin，SHAN XiuMing，R EN Yong1.Fractional Fourier domain analysis of decimation and interpolat ion MENG XiangYi，TAO Ran，WANG Yue2.A reduced state SISO iterative decoding algorithm for serially concatenated continuous phase modulation SUN JinHua，LI JianDong，JIN LiJun3.On the linear span of the p-ary cascaded GMW sequences TA NG XiaoHu4.De-interlacing technique based on total variation with spatial-t emporal smoothness constraint YIN XueMin，YUAN JianHua，LU Xia oPeng，ZOU MouYan5.Constrained total least squares algorithm for passive location based on bearing-only measurements WANG Ding，ZHANG Li，WU Ying6.Phase noise analysis of oscillators with Sylvester representation for periodic time-varying modulus matrix by regular perturbations FAN JianXing，YANG HuaZhong，WANG Hui，YAN XiaoLang，HOU ChaoHuan7.New optimal algorithm of data association for multi-passive-se nsor location system ZHOU Li，HE You，ZHANG WeiHua8.Application research on the chaos synchronization self-mainten ance characteristic to secret communication WU DanHui，ZHAO Che nFei，ZHANG YuJie9.The changes on synchronizing ability of coupled networks fro m ring networks to chain networks HAN XiuPing，LU JunAn10.A new approach to consensus problems in discrete-time mult iagent systems with time-delays WANG Long，XIAO Feng11.Unified stabilizing controller synthesis approach for discrete-ti me intelligent systems with time delays by dynamic output feedbac k LIU MeiQin1.Survey of information security SHEN ChangXiang，ZHANG Hua ngGuo，FENG DengGuo，CAO ZhenFu，HUANG JiWu2.Analysis of affinely equivalent Boolean functions MENG QingSh u，ZHANG HuanGuo，YANG Min，WANG ZhangYi3.Boolean functions of an odd number of variables with maximu m algebraic immunity LI Na，QI WenFeng4.Pirate decoder for the broadcast encryption schemes from Cry pto 2005 WENG Jian，LIU ShengLi，CHEN KeFei5.Symmetric-key cryptosystem with DNA technology LU MingXin，LAI XueJia，XIAO GuoZhen，QIN Lei6.A chaos-based image encryption algorithm using alternate stru cture ZHANG YiWei，WANG YuMin，SHEN XuBang7.Impossible differential cryptanalysis of advanced encryption sta ndard CHEN Jie，HU YuPu，ZHANG YueYu8.Classification and counting on multi-continued fractions and its application to multi-sequences DAI ZongDuo，FENG XiuTao9.A trinomial type of σ-LFSR oriented toward software implemen tation ZENG Guang，HE KaiCheng，HAN WenBao10.Identity-based signature scheme based on quadratic residues CHAI ZhenChuan，CAO ZhenFu，DONG XiaoLei11.Modular approach to the design and analysis of password-ba sed security protocols FENG DengGuo，CHEN WeiDong12.Design of secure operating systems with high security levels QING SiHan，SHEN ChangXiang13.A formal model for access control with supporting spatial co ntext ZHANG Hong，HE YePing，SHI ZhiGuo14.Universally composable anonymous Hash certification model ZHANG Fan，MA JianFeng，SangJae MOON15.Trusted dynamic level scheduling based on Bayes trust model WANG Wei，ZENG GuoSun16.Log-scaling magnitude modulated watermarking scheme LING HeFei，YUAN WuGang，ZOU FuHao，LU ZhengDing17.A digital authentication watermarking scheme for JPEG image s with superior localization and security YU Miao，HE HongJie，ZHA NG JiaShu18.Blind reconnaissance of the pseudo-random sequence in DS/ SS signal with negative SNR HUANG XianGao，HUANG Wei，WANG Chao，L(U) ZeJun，HU YanHua1.Analysis of security protocols based on challenge-response LU O JunZhou，YANG Ming2.Notes on automata theory based on quantum logic QIU Dao Wen3.Optimality analysis of one-step OOSM filtering algorithms in t arget tracking ZHOU WenHui，LI Lin，CHEN GuoHai，YU AnXi4.A general approach to attribute reduction in rough set theory ZHANG WenXiuiu，QIU GuoFang，WU WeiZhi5.Multiscale stochastic hierarchical image segmentation by spectr al clustering LI XiaoBin，TIAN Zheng6.Energy-based adaptive orthogonal FRIT and its application in i mage denoising LIU YunXia，PENG YuHua，QU HuaiJing，YiN Yong7.Remote sensing image fusion based on Bayesian linear estimat ion GE ZhiRong，WANG Bin，ZHANG LiMing8.Fiber soliton-form 3R regenerator and its performance analysis ZHU Bo，YANG XiangLin9.Study on relationships of electromagnetic band structures and left/right handed structures GAO Chu，CHEN ZhiNing，WANG YunY i，YANG Ning10.Study on joint Bayesian model selection and parameter estim ation method of GTD model SHI ZhiGuang，ZHOU JianXiong，ZHAO HongZhong，FU Qiang。

cv人物表

CV人物4：Matthew Turk毕业于MIT，最有影响力的研究成果：人脸识别。其和Alex Pentland在1991年发表了"Eigenfaces for Face Recognition".该论文首次将PCA（Principal Component Analysis）引入到人脸识别中，是人脸识别最早期最经典的方法，且被人实现，开源在OpenCV了。主页：/~mturk/
CV人物16： William T.Freeman毕业于MIT；研究领域：应用于CV的ML、可视化感知的贝叶斯模型、计算摄影学；最有影响力的研究成果：图像纹理合成；Alex Efros和Freeman在2001年SIGGRAPH上发表了"Image quilting for texture synthesis and transfer"，其思想是从已知图像中获得小块，然后将这些小块拼接mosaic一起，形成新的图像。该算法是图像纹理合成中经典中的经典。主页：/billf/
CV人物17： Feifei Li李菲菲，毕业于Caltech；导师：Pietro Perona；研究领域：Object Bank、Scene Classification、ImageNet等；最有影响力的研究成果：图像识别；她建立了图像识别领域的标准测试库Caltech101/256。是词包方法的推动者。主页：/~feifeili/
CV人物8：Michal Irani毕业于Hebrew大学，最有影响力的研究成果：超分辨率。她和Peleg于1991年在Graphical Models and Image Processing发表了"Improving resolution by image registration",提出了用迭代的、反向投影的方法来解决图像放大的问题，是图像超分辨率最经典的算法。我在公司实现的产品化清晰化增强算法就参考了该算法思想哈哈。主页：http://www.wisdom.weizmann.ac.il/~irani/

笛卡尔的本体论之争

笛卡尔的本体论之争首先周一公布2001年6月18日;实质性修改太阳2006年10月15日笛卡尔的本体论（或先验）的论点，既是哲学的一个最迷人，他的理解方面的不足。

论据与魅力源于努力证明神的存在，从简单的处所，但功能强大。

存在是产生立即从清晰和明确的想法是一个无比完美。

讽刺的是，简单的说法也产生了一些误读，加剧了部分由笛卡尔没有一套单一版本。

该声明的论点主要出现在第五沉思。

这种说法因果来得早在接踵而至的一个神的存在，沉思在第三，不同的证据提出问题的两项之间的秩序和关系。

重复笛卡尔哲学原理，包括本体论争论的几个文本等中央。

他还辩解首先由一些主要的知识分子，他在一天，严厉打击反对第二次回复，和第五。

笛卡尔不是第一位哲学家，制订一个本体论的论点。

一个早期版本的说法已大力安瑟伦辩护圣在11世纪，然后圣托马斯阿奎那批评由当代），后来被命名为Gaunilo和尚（安瑟伦（尽管他的言论是针对然而，另一个版本参数）。

阿奎那的批评被视为如此具有破坏性，本体论的争论了数百年死亡。

它的出现，作为一个同时代的惊喜笛卡尔，他应该试图复活它。

虽然他声称没有被证明的熟悉安瑟伦的版本，笛卡尔似乎他自己的工艺参数，以阻止传统的反对。

尽管相似之处，笛卡尔的论点的版本不同于安瑟伦方式在重要的。

后者的版本被认为要从定义这个词的含义“上帝”，上帝是一个被一大于不能设想。

笛卡尔的观点相反，中，主要是基于两个他的哲学的中心原则-天生的思想理论和学说明确的印象和独特的。

他声称不依赖于上帝的任意定义，而是一种天生的想法，其内容是“的。

” 笛卡尔的版本也非常简单。

神的存在是直接从推断的事实，有必要存在的想法是包含在一个清晰而鲜明的超级完美的存在。

事实上，在一些场合，他建议，所谓的本体论“的论调”是不是一个正式的哲学偏见的证据，而是在所有不言而喻的公理直观地掌握了一个心灵的自由。

笛卡尔的本体论的争论相比往往以几何论证，认为有必要存在的想法不能排除再从神比事实平等的角度，其角度，例如两权，可以被排除在一个三角形的想法。

基于语谱图的声乐分析

技术创新《微计算机信息》(管控一体化)2010年第26卷第7-3期360元/年邮局订阅号：82-946《现场总线技术应用200例》博士论坛基于语谱图的声乐分析Analysis of vocality on Spectrogram(上海理工大学)陈青龚乾张鸣CHEN Qing GONG Qian ZHANG Ming摘要:语谱图通常从时域信号中用短时傅里叶变换计算得到,并表示成灰度图像。

为了得到更高的分辨率和更好的视觉效果来进行语音信号的处理与分析,本文利用MATLAB 的编程算法和函数库,对运算得到的语谱图使用伪彩色映射算法来观察能量分布,并对声音作进一步谐波跟踪研究,做到对不同素材的音乐性判定。

关键词:时频分析;语谱图;伪彩色映射;谐波跟踪中图分类号:TP391文献标识码:AAbstract:Spectrogram is usually calculated from the time-domain signal by short-time Fourier transform and displayed in gray scale images.In this paper,with the programming and library of MATLAB,the spectrums of speech energy distributions are mapped to pseudo-color images in order to get higher resolution and better visual effects for speech signal processing and some analysis for har -monic tracking are used to these signals from music.At last,the conclusion can contact to musical identification.Key words:Time-frequency Analysis;Spectrogram;Pseudo-color;Harmonic track文章编号:1008-0570(2010)07-3-0006-03引言语谱图是一种在语音分析以及语音合成中具有重要实用价值的时频图,它可以反映语音信号动态频谱特性,被视为语音信号的可视语言。

Classifying heart sounds using peak location for segmentation and feature construction

Classifying heart sounds using peak location for segmentation andfeature constructionElsa Ferreira Gomes Emanuel PereiraGECAD-Knowledge Eng.Decision Support Institute of Engineering(ISEP/IPP)Porto,Portugal Institute of Engineering(ISEP/IPP)Porto,PortugalAbstractIn this paper we describe our methodologyand results for the Classifying Heart SoundsPASCAL Challenge.We present our algo-ritm and results for segmentation and classi-ﬁcation of S1and S2heart sounds.1IntroductionThe goal of this challenge is to produce methods for Classifying Heart Sounds PASCAL Challenge.Specif-ically,the aim of this work is to create theﬁrst level of screening of cardiac pathologies both in a Hospi-tal environment by a doctor(using a digital stetho-scope)and at home by the patient(using a mobile de-vice).The main components of heart sound signal of a normal heart are theﬁrst heart sound,S1(or lub), corresponding to the systolic period,and the second heart sound,S2(or dub),the diastolic period(Gupta et.al,2007).In fact,this challenge is composed by two diﬀerent challenges,Challenge1and Challenge2.The aim of Challenge1,Heart Sound Segmentation,is to produce a method that can locate S1and S2sounds within audio data,segmenting the Normal audioﬁles for two datasets,Datataset A and Dataset B.Dataset A contains data from the general public via the iS-tethoscope Pro iPhone app.Dataset B,contains data from a clinic trial in hospitals using the digital stetho-scope DigiScope(Bentley et.al,2011).These audio ﬁles are of varying lengths,between1second and30 seconds(some have been clipped to reduce excessive noise and provide the salient fragment of the sound). The aim of Challenge2,Heart Sound Classiﬁcation,is to produce a method that can classify real heart au-dio or beat classiﬁcation into one of four categories for Preliminary work.Under review by AISTATS2012.Do not distribute.Dataset A(Normal,Murmur,Extra Heart Sound and Artifact)and three categories for Dataset B(Normal, Murmur and Extrasystole).Data is gathered in real-world situations and frequently contains background noise of every conceivable type.The diﬀerences be-tween heart sounds corresponding to diﬀerent heart symptoms can also be extremely subtle and challeng-ing to separate.Success in classifying this form of data requires extremely robust classiﬁers.Despite its med-ical signiﬁcance,to date this is a relatively unexplored application for machine learning(Bentley et.al,2011). Here we present our work developed to answer to the proposed challenges,Challenge1and Challenge2.2The challengesAs we said in the previous section,this challenge is, in fact,composed by two challenges.The goal of the ﬁrst one is the heart sound segmentation and the goal of the second one is the heart sound classiﬁcation.We know data have been gathered from two sources,re-sulting in two sets:Dataset A and Dataset B.We have diﬀerent categories of data for each dataset.In the Normal category there are normal,healthy heart sounds.However,these sounds may contain a vari-ety of background noises.They may also contain oc-casional random noise corresponding to breathing,or brushing the microphone against clothing or skin.A normal heart sound has a clear lub dub,lub dub pat-tern,with the time from lub to dub shorter than the time from dub to the next lub(when the heart rate is less than140beats per minute).We know most normal heart rates at rest are between about60and 100beats(lub and dub)per minute.However,in these data,since the data may have been collected from children or adults in calm or excited states,the heart rates in the data may vary from40to140beats or higher per minute.In the Murmur category,the heart murmurs sound as though there is a whoosh-ing,roaring,rumbling,or turbulentﬂuid noise in one of two temporal locations:(1)between lub and dub,or(2)between dub and lub.They can be a symp-tom of many heart disorders,some serious.One of the things that confuses non-medically trained people is that murmurs happen between lub and dub or be-tween dub and lub;not on lub and not on dub.We also have the Extra Heart Sound category but only for Dataset A.These Extra heart sounds can be identiﬁed because there is an additional sound,e.g.a lub-lub dub or a lub dub-dub.An extra heart sound may not be a sign of disease.However,in some situations it is an important sign of disease,which if detected early could help a person.The extra heart sound is impor-tant to be able to detect as it cannot be detected by ultrasound very well.In the Artifact category,only existing on Dataset A,there are a wide range of dif-ferent sounds,including feedback squeals and echoes, speech,music and noise.There are usually no dis-cernible heart sounds,and thus little or no temporal periodicity at frequencies below195Hz.This category is the most diﬀerent from the others.It is important to be able to distinguish this category from the other three categories,so that someone gathering the data can be instructed to try again.In the Extrasystole category,only existing in Dataset B,sounds may ap-pear occasionally and can be identiﬁed because there is a heart sound that is out of rhythm involving extra or skipped heartbeats,e.g.a lub-lub dub or a lub dub-dub.Notice that an extrasystole may not be a sign of disease.It can happen normally in an adult and can be very common in children.However,in some situa-tions extrasystoles can be caused by heart diseases.So, theﬁles Atraining normal.zip,Atraining murmur.zip, Atraining extrahs.zip,Atraining artifact.zip,Aunla-belledtest.zipﬁles were available for DatasetA.For Dataset B we had Btraining normal.zip,Btrain-ing murmur.zip,Btraining extrasystole.zip and Bun-labelledtest.zip.3Challenge1-Heart SoundSegmentationIn theﬁrst challenge we aim to produce a method for determining the location of S1and S2sounds within audio data,segmenting the Normal audioﬁles in Dataset A and Dataset B.The recorded signals were ﬁrst preprocessed before performing segmentation.In theﬁrst step the signals were down sampled andﬁl-tered.In the second step,we have done the segmen-tation.Our algorithm is based on the envelope cal-culated using the normalized average Shannon energy (Liang,1997).3.1Pre-processingThe original signal was decimated,using the decimate function of Matlab(MATLAB,2009b)with factor5. Then,a band-passﬁlter was applied.Considering the frequency components of S1and S2heart sounds,the chosenﬁlter was aﬁfth order Chebyshev type I lowpass ﬁlter with cutoﬀfrom100Hz to882Hz.Then,the signals were normalized to the absolute maximum of the signal(Liang,1997).3.2SegmentationAfter pre-processing,we calculated the Shannon En-velope of the normalized signal.Then,the Average Shannon Energy is calculated in continuous0.02sec-onds windows throughout the signal with0.01second overlapping(Liang,1997).3.2.1PeakﬁndingAfter obtaining the normalized average Shannon en-ergy curve we identiﬁed the peaks.For that,we worked on the basis of the open source function peakdet,written in Matlab,with some changes made by us.This functionﬁnds the local maxima(and min-ima)using the strategy that a point is considered a maximum peak if it has a locally maximal value,and was preceded(to the left)by a value lower than a given delta.We have used two parameters of the function. Theﬁrst is the vector to examine,and the second is the peak gap threshold(the delta).In our case con-sidering this delta on the y values it was not enough. We also had to control the distance between peaks on the x-axis,because we know we cannot have two heart sounds too close.So,we have changed the function so that a local maximum is considered a peak if the distance to the nearest a peak is greater than a sec-ond ing this,we segmented almost all heart sounds.However,we also need to distinguish between S1and S2.Our current approach for S1/S2 discrimination is still unsatisfactory.First,we tried to perform the detection of S1and S2sounds based on the fact that S2is longer than S1,for normal heart rates(Kumar,2006).Bearing this in mind we tried to pick each heart cycle and the corresponding systolic in-terval.The duration of S1-S2sounds,or the distance between S1and S2,was calculated and compared for every segment(Gupta et.al,2007).The longest inter-val between two sounds was considered to correspond to the diastolic period and the sound at the right side was assigned as S1and the sound at the left side was assigned as S2.Unfortunately,weﬁnd that those in-tervals vary widely fromﬁle toﬁle,in our datasets. This happens because there are very diﬀerent kinds of heart sounds data,for both datasets.We also tried touse a similar process to the energy,assuming S2has higher frequency,but until now,we had no success.We are working on it.3.3ResultsAs we said before the task of the ﬁrst challenge is to segment the audio ﬁles in Atraining normal.zip and Btraining normal.zip (for Dataset A and Dataset B,respectively).Our results were evaluated on a pro-vided validation set with the correct locations of S1and S2sounds.This set contains the segmentation for sounds of the normal category from Dataset A and Dataset B.A test set for ﬁnal evaluation was also avail-able.This set contained hidden locations but provided the ﬁnal evaluation results.In Figure 1we can see the result of our method for the peak detection for ﬁle 1031305031931979D2.wav.In the 1st chart we have the original signal.In the 2nd chart we have the decimated signal.In the 3rd chart we have decimated signal ﬁltered with a Chebyshev ﬁlter.In the 4rth chart we have the envelope of the signal and peaks.In the 5th chart we have the peaks over the originalsignal.Figure 1:Peak detection on heart sound signal In Table 1we present our results for Dataset B.As we can see we have an error of 18.1for ﬁle 1751307987962616B1but an error 24149.9for ﬁle 1261306777102824B.The total error was 72242.8.Some of these results could be better if we choose the ﬁrst value for S1or S2.However,because our method is not correctly implemented for the detection of S1and s2sounds,we have placed all ﬁrst occurrences as S1.For example,for ﬁle 1541306935608852B1if we start with a S2the error decreases to 51.3instead the error 2139.9appearing in the table.For Dataset A the error is higher then the error ob-Table 1:Results for Dataset B File Name Heart Avgbeats Error 1031305031931979B 12.554.31031305031931979D210.035.71061306776721273B1 4.035.11061306776721273C2 3.027.31061306776721273D1 3.5121.91061306776721273D28.04084.81071305654946865C17.51545.61261306777102824B 5.024149.91261306777102824C 3.013871.31331306759619127A 4.51578.21341306428161797C2 2.545.11371306764999211C 15.01629.01401306519735121B 11.049.41461306778707532B 18.02121.91461306778707532D3 3.026.81471306523973811A 3.53095.81481306768801551D2 5.010226.91511306779785624D 4.52560.31541306935608852B1 4.52139.91591307018640315B1 6.078.01591307018640315B2 3.066.71671307111318050A 13.058.01671307111318050C 3.51484.81721307971284351B1 3.568.41751307987962616B1 2.518.11751307987962616D 7.51813.21791307990076841B 16.563.21811308052613891D 3.040.01841308073010307D 26.570.71901308076920011D4.01082.6tained for dataset B (see Table 2).4Challenge2-Heart Sound ClassiﬁcationThe task of Challenge 2is to produce a method that can classify the real heart audio into one of four cat-egories for Dataset A (Normal,Murmur,Extra Heart Sound and Artifact)and three classes for Dataset B (Normal,Murmur and Extrasystole).This phase in-volves feature construction and selection and the goal of this phase of the challenge is correctly labeling the sounds in the provided ﬁles Aunlabelledtest.zip and Bunlabelledtest.zipTable2:Results for Dataset AFile Name Heart Avgbeats Error201101070538.aif11.515324.8201101151127.aif 1.0516698.0201102081152.aif7.5156064.2201102201230.aif9.588952.8201102270940.aif 1.01445703.5201103101140.aif 5.0374151.6201103140135.aif 5.0314072.2201103170121.aif 5.5299314.2201104122156.aif 2.0640841.8201106151236.aif 5.0368613.54.1Feature construction and selectionAfter the pre-processing and segmentation of the heart sound signal,some features were extracted.Currently, we are using six features(3),four of them were ex-tracted from the distance between S1and S2.Assum-ing that S1corresponds to smaller segments and S2to the others,we consider the ratio of the standard devia-tion of S1over whole standard deviation and S2(Rs1 and Rs2,respectively)and the ratio of the mean of S1over the total mean and the ratio of S2(Rm1and Rm2,respectively).Another feature,Rmedian,is the ratio of the median of the three largest segments in the sample over total mean.The sixth feature,R2,is r square of the array of the sorted segments of the sam-ple.After obtaining the features we used two diﬀerent methods from the Weka data mining suite(Witten and Frank,2005):J48,which generates decision trees,and MLP,the Multi Layer Perceptron.Table3:FeaturesFeature DescriptionRs1σS1/σtotalRs2σS2/σtotalRm1µS1/µtotalRm2µS2/µtotalRmedian median(max1,max2,max3)/µtotalR2r square of the sorted segments4.2ResultsIn Challenge2,the eﬀectiveness of our classiﬁcation approach is assessed using three metrics(per dataset) which are calculated from the tp(true positives),fp (false positives),tn(true negaties)and fn(false nega-tives)values.The metrics are precision per class,the Youden’s Index,the F-score(only for Dataset A)and the Discriminant Power(only for Dataset B).Preci-sion gives us the positive predictive value(the pro-portion of samples that belong in category c that are correctly placed in category c).Youden’s Index has been used to compare diagnostic abilities of two tests, by evaluating the algorithm’s ability to avoid failure. In Dataset A,Youden’s Index is evaluated for Artifact category.In Dataset B the Youden’s Index is calcu-lated for the problematic heartbeats(Murmur and Ex-trasystole categories combined).Discriminant Power evaluates how well an algorithm distinguishes between positive and negative examples.It is a poor discrim-inant for a value¡1,limited for a value¡2,fair for a value¡3,and good in other cases.Here we calcu-late the Discriminant power of problematic heartbeats (Murmur and Extrasystole categories combined).In Table4and Table5,we present the results for Dataset A and Dataset B,obtained after applying the J48and MLP methods.As in Challenge1,a testset was provided where we could test our method’s eﬀec-tiveness on the unlabelled set.Table4:Challenge2evaluation for DatasetAJ48MLP Precision of Normal0.250.35 Precision of Murmur0.470.67 Precision of Extrasound0.270.18 Precision of Artifact0.710.92 Artifact Sensitivity0.630.69 Artifact Speciﬁcity0.390.44 Heartproblem Detection Sensitivity0.550.45 Heartproblem Detection Precision0.400.43 Youden Index of Artifact0.010.13 F-Score of Heartproblem Detection0.200.20 Total Precision 1.71 2.12 Table5:Challenge2evaluation for DatasetBJ48MLP Precision of Normal0.720.70 Precision of Murmur0.320.30 Precision of Extrasystole0.330.67 Sensitivity of heart problem0.220.19 Speciﬁcity of heart problem0.820.84 Youden Ind Heartproblem Detection0.040.02 Discriminant Power0.050.04 Total Precision 1.37 1.67As we can see in Table5,our method has problemsin classifying the non-normal heart beats,for Dataset B.In Dataset A the situation is diﬀerent.In this case the normal class is one of the most diﬃcult(Table 4).However,we think we can improve our method by improving the correct identiﬁcation of S1and S2 in the segmentation and byﬁnding new features that take advantage of this identiﬁcation.5ConclusionsIn this paper,we present a methodology for the Clas-sifying Heart Sounds PASCAL Challenge.We pro-pose an algorithm for S1and S2heart sound identi-ﬁcation(without ECG reference).The segmentation is accomplished by using the envelope of Shannon en-ergy and an algorithm for peak detection.Despite of the good performance for the correct detection of S1 and S2sounds in the signal,we need to improve the criteria for identifying S1and S2(who is who).After the segmentation,we used J48and MLP algorithms (using Weka)to train and classify the computed fea-tures into Normal,Murmur or Extrasystole for Dataset B and Normal,Murmur,Extrasound and Artifact for Dataset A.We think this is a good basis for further analysis of the heart sound signals.ReferencesH.Liang,S.Lukkarinen,and I.Hartimo.(1997) Heart Sound Segmentation Algorithm Based On Heart Sound puters in Cardiology24:105-108.P.Bentley,G.Nordehn,M.Coimbra,and S.Mannor.The PASCAL Classifying Heart Sounds Challenge2011(CHSC2011)Results, /heartchallenge/index.html.D.Kumar,P.Carvalho,M.Antunes,P.Gil,J. Henriques,and L.Eug.(2006)A new algorithm for detection of s1and s2heart sound.IEEE International Conference on ICASSP2006,pp.1180-1183.C.N.Gupta,R.Palaniappan,S.Rajan,S.Swami-nathan,and S.M.Krishnan(2005)Segmentation and classiﬁcation of heart sounds1.Canadian Conference on Electrical and Computer Engineering.1674-1677.C.Gupta,R.Palaniappan,S.Swaminathan,and S.Krishnan.(2007)Neural network classiﬁcation of homomorphic segmented heart sounds.Applied Soft Computing7(1):286-297.E.Billauer,3.4.05(2011),www.billauer.co.il/ peakdet.htmlMATLAB version7.9.0.Natick,Massachusetts.The MathWorks Inc.,2009b.I.H.Witten and E.Frank(2005)Data Mining:Prac-tical machine learning tools and techniques,2nd Edi-tion,Morgan Kaufmann,San Francisco,2005.。

routine练习题

routine练习题一、词汇练习1. 选择正确的单词填空：1. I usually _______ to work bus.2. She _______ her homework every evening.A. doesB. doC. does not doD. doesn't do3. They _______ a movie last night.A. watchB. watchesC. watchedD. watching2. 选择正确的词组：1. I _______ (go, going) to the gym this morning.2. He _______ (be, is) late for school again.3. She _______ (do, does) her homework every day.3. 选择正确的形容词：1. This is a _______ (good, bad) book.2. She is a _______ (smart, silly) girl.3. The weather is very _______ (hot, cold) today.二、语法练习1. 选择正确的时态：1. I _______ (go, went) to the park yesterday.2. She _______ (be, was) happy when she received the gift.3. They _______ (do, did) their homework last night.2. 选择正确的语态：1. The teacher _______ (teach, is taught) Mr. Wang.2. The book _______ (write, is written) a famous author.3. The letter _______ (send, is sent) to her last week.3. 选择正确的连词：1. I _______ (go, am going) to the movies, _______ (because, because of) I have free time.2. She _______ (like, likes) coffee, _______ (but, but) she doesn't like tea.3. I _______ (finish, finished) my homework, _______ (so, so) I can go out now.三、阅读理解1. 阅读短文，回答问题：1. What is the main idea of the passage?2. Who is the main character in the story?3. What happens at the end of the passage?2. 阅读文章，判断正误：1. The story is about a boy who goes to the park every weekend.2. The boy meets his friends at the park and they play games together.3. The boy goes home after playing games with his friends.3. 阅读文章，找出关键信息：1. What is the author's favorite color?2. Why does the author like this color?3. What does the author think about other colors?四、写作练习1. 介绍动物的名字和种类。

1、下载文档前请自行甄别文档内容的完整性，平台不提供额外的编辑、内容补充、找答案等附加服务。
2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
3、如文档侵犯您的权益，请联系客服反馈,我们会尽快为您处理(人工客服工作时间：9:00-18:30)。

Stochastic Human Segmentation from a Static Camera Tao Zhao Ram NevatiaUniversity of Southern CaliforniaInstitute for Robotics and Intelligent SystemsLos Angeles,CA90089-0273taozhao nevatia@AbstractSegmenting individual humans in a high-density scene(e.g., a crowd)acquired from a static camera is challenging mainlydue to object inter-occlusion(Fig.1).We deﬁne this problem asa“model-based segmentation”problem and the solution is ob-tained using a Markov chain Monte Carlo(MCMC)approach.Knowledge of various aspects including human shape,humanheight,camera model,and image cues including human head candidates,foreground/background separation are integratedin a Bayesian framework.We show promising results on some challenging data.1Introduction and MotivationSegmentation and tracking of humans in video sequences is important for a number of tasks such as video surveillanceand event inference as humans are the principal actors in dailyactivities of interest.We consider scenarios where the camera isﬁxed;in this case,moving pixels(foreground)can be de-tected fairly reliably by comparing the incoming frames with a background model.However,the problem of obtaining object level descriptionsfrom moving pixels begins to gain more attention.The difﬁ-culty of this problem increases with the density of the objectsin the scene.We would like to categorize the characteristics ofthe problem according to the object density as follows: Low density:The moving objects are sparse and seldom overlapped.In this case,connected components of theforeground(called blob)correspond well to individualobjects.Moderate density:Moving objects have simple interac-tions such as passing-by or small groups of people.In this case,using the knowledge of human head position and unexplained foreground([11][7])or vertical projec-tion of the blob([4])is able to solve the segmentation problem satisfactorily in most cases.Besides,a human speciﬁc model initialized before occlusion can also help solve the problem([3][11]).1This research was supported,in part,by the Advanced Research and De-velopment Activity of the ernment under contract No.MDA-908-00-C-0036.(a)(b)Figure1:A sample input frame(a)and its foreground from standard background subtraction(b).High density:Blobs consist of large groups of people.People in the scene occlude each other either due to their spatial proximity or due to camera projection(an exam-ple frame and its foreground are shown in Fig.1).In this case,many of the heads are not on the foreground bound-ary and the vertical projection of the big blob is also not informative enough to perform segmentation.Previous research on segmenting and tracking of multiple humans([8][4][6][3][11][7],etc)has been focused mainly on theﬁrst two classes.The high density case is interesting because such scenes usually contain rich human behaviors of interests.However the challenge is also obvious.Color seg-mentation is not likely to segment the individual humans;mo-tion segmentation also won’t give satisfactory result due to the similarity of the motion of individuals in a group and the non-rigid motion of human.Face detection may not be effective due to the arbitrary viewpoint(consider viewed from back) and the low image resolution.1.1Problem deﬁnitionOur goal is to segment overlapping humans from the fore-ground.We deﬁne the problem as a model-based segmentation problem.The difference from general image(region)segmen-tation is that we use object shape model to constrain the seg-mentation instead of enforcing some homogeneity(e.g.,color, texture,etc)within the segments.The difference from object detection methods is that we know in advance that the image region of interest is made up of only the objects of known classes(or largely so)but the objects may occlude each other(severely).Assume we have an object model M and an image region (foreground)is made up of an unknown number()of suchobjects with parameters,where may con-tain position,size and other shape parameters.A solution is in the form of.Such a joint state has to be used due to their possible overlapping.Obviously,the solutionspace is of unknown dimensions.The problem is toﬁnd a solution which maximize a posteriori probability(MAP):(1)where is various features that we can extract from the image.In other words,we need toﬁnd the number of humans and the location,size and other parameters for each human which maximize the posterior probability.The posterior probability is further decomposed into a likelihood term and a prior term according to Bayes rule:(2) where is a normalization factor independent to.1.2Outline of our approachOptimization in such a complex solution space is generally difﬁcult.The Markov chain Monte Carlo(MCMC)provides a sampling method to traverse the complex solution space with a Markov chain and the optima can be found during the traver-sal.Recently,DDMCMC(data-driven MCMC)has been pro-posed in[12]and[10]for object recognition and image seg-mentation which emphasizes incorporating domain knowledge (heuristics)to design the proposal probability of the Markov chain to make the traversal more efﬁcient compared to the tra-ditional more random proposal probability.Our work is also motivated by this idea.A block diagram of our system is shown in Fig.2.The foreground is computed by subtracting the background from the input image frames.The foreground is to be segmented into different overlapping human objects by the MCMC ap-proach.We use a human model consisting of three parts to capture the gross shape of standing and walking humans.A prior based on the number of humans makes the segmenta-tion concise(Sec.3.1).A joint likelihood based on the fore-ground/background separation is deﬁned to minimize the dif-ference of the real foreground and the foreground correspond-ing to the solution(Sec.3.4).Observing the head might be the only feature in the presence of severe inter-occlusion, we describe two techniques to detect heads based on the fore-ground boundary and intensity edges respectively(Sec.3.2). This is used as domain knowledge to direct the Markov chain on where to create new human hypotheses.Together with re-moval and diffusion processes,a well-balanced Markov chain dynamics is designed(Sec.3.3).We show promising experi-mental results on both outdoor and indoor sequences(Sec.4).Figure2:The block diagram of the system.Shaded box:pro-gram module;plane box:model;thick arrow:dataﬂow;thin arrow:model association.2The MCMC Framework for MAP Estima-tionSuppose we want to compute the maximum a posteriori value from a probability distribution(we usefor clarity),this is done by sampling fromandﬁnding the maximum value.To sample from,a Markov chain can be designed.A widely used algorithm is the Metropolis-Hasting algorithm.At each iteration,we sample a candidate state from a proposal distribution(in simple words,what new state should the Markov chain go to from the current state.).The candidate state is accepted with the following probability:If the candidate state is accepted,,otherwise, .It can be proven that the Markov chain constructed this way always has its stationary distribution equal to, independent of the choice of the proposal probability and the initial state[9].The advantage of the MCMC approach over other optimization methods is that with jump and diffu-sion dynamics it can work in complex solution space such as the one here with varying number of dimensions.However,the choice of the proposal probability can affect the efﬁciency of the MCMC signiﬁcantly.A random proposal probability will lead to very slow convergence rate while a proposal probability designed with domain knowledge ([12][10])will make the Markov chain traverse the solution space more efﬁciently.If the proposal probability is informa-tive enough so that each sample can be thought of as a hypoth-esis,then the MCMC approach can be thought of as a stochas-tic version of the widely used hypothesis and testing approach ([12]).In our problem,head candidates which are detected from both the foreground boundaries and the intensity edges serve as domain knowledge on where to generate human hypotheses.(a)(b)Figure3:3-ellipsoid human model.(a)The model is com-posed of3parts:head,torso and legs;(b)The2D heightscomputed by camera model from aﬁxed head-top assumingthe3D heights are1.6,1.7and1.8meters respectively.The components of the MCMC framework are described indetail in the next section.3Human Segmentation from Foreground Us-ing MCMC3.1Human model and human count priorHuman body is highly articulated.To model it precisely,acomplicated kinematics model with over20DOF is needed.However,in our application the human image is generallysmall and human motion is mostly limited to standing or walk-ing.We can thus use a low dimensional model which is acomposition of3ﬁxed ellipsoids:one for the head,one for thetorso and one for the legs(Fig.3.(a)).An ellipsoidﬁts theseforms well and has a nice property that its projection is an el-lipse with a convenient form[5].The parameters of the humanmodel include position,height and fatness().The height parameter also affects the width of the human shapeproportionally and the fatness parameter captures the extra fat-ness besides the proportional change.We assume that humanheight has a uniform distribution in the range ofand the fatness is in the range of times of the averageone.We also assume that the camera model and the groundplane are known.The camera model along with the heightrange can add constraints to the image height of a human.Fig.3.(b)shows the projected human model computed by thecamera model with the2D head positionﬁxed assuming the3D heights are1.6,1.7and1.8meters respectively.A prior on the number of humans in the scene is applied asin Equ.2to make the segmentation concise by penalizingredundant human hypotheses which do not contribute enoughto the joint likelihood.(3)where is a coefﬁcient whose value is only meaningful rel-ative to the coefﬁcient in the likelihood measurement.There-fore,we set.3.2Head candidate detectionWe observe that the head is almost the only reliably visiblefeature for a human in highly occluded environments observedfrom above.Therefore,we developed two techniques to de-tect head from image frames based on foreground boundaryand the intensity edges respectively.The head candidates set.3.2.1Head candidates from foreground boundary()This method detects the heads which are on the boundaryof the foreground[11].The basic idea is toﬁnd the local peaksof the boundary.The peaks are furtherﬁltered by checking ifits corresponding human is higher than a minimum interestedheight and if there are enough foreground pixels below it ac-cording to a human height range and the camera model.Thisdetector has a high detection rate and is also effective when thehuman is small and image edges are not reliable;however,itcannot detect the heads in the interior of the foreground blob.Fig.4.(a)shows the on the example frame.3.2.2Head candidates from intensity edges()Here we describe a head detector based on image intensityedges which is also effective for the heads in the interior of theblob.First Canny edge detection[2]is applied to the dilated fore-ground region of the input image.A distance transformation[1]is computed on the edge map.Fig.4.(b)shows the expo-nential edge map where(is the distance to the closest edge point and is a factor tocontrol the responseﬁeld and set to.).Besides,the coor-dinates of the closest pixel point are also recorded as.The unit image gradient vector is only computed atedge pixels.The“”shape of head and shoulder contour(Fig.4.(c))iseasily derived from the3-ellipsoid human model.The head-shoulder contour is generated from the projected ellipses bytaking the whole head and the upper quarter torso as the shoul-der.The normals of the contour points are also computed.Thesize of the human model is determined by the camera calibra-tion assuming a known height(1.6,1.7,1.8meters are usedand the maximum response is recorded).Denote and as the positions andthe unit normals of the model points respectively when headtop is at.The model is matched with the image in thefollowing way.A head candidate map is constructed by evaluatingon every pixel in the dilated foreground region.The detectionscore is alsoﬁltered in the same way as described in Sec.3.2.1.After smoothing it,weﬁnd all the peaks above a thresholdwhich ensures an almost detection rate but results in ahigh false alarm rate.An example is shown in Fig.4.(d).The(a)(b)(c)(d)Figure4:Head detectors.(a)Head detection from mo-tion boundaries();(b)Distance transformation onCanny edge detection result;(c)The head-shoulder()model:red-head part,blue-shoulder part,green-normals;(d)Head de-tection from intensity edges().See Sec.3.2for detail.false alarms tend to happen in the area of rich texture wherethere are abundant edges of various orientations.3.3Dynamics of the Markov chainDenote the state at iteration as.The following Markov chain dynam-ics are designed which corresponds to sampling the proposalprobability:Add a new human hypothesis.Randomly select a headcandidate which is not too close to anycurrent heads.A new human hypothesis is assem-bled as where heightand fatness are randomly assigned..Remove an existing human hypothesis.Randomly se-lect an existing human to remove..Change the parameters of a human hypothesis.Ran-domly select an existing human and add a randomnoise to its parameters.,.Theﬁrst two are referred to as jump dynamics and the lastone is referred to as diffusion dynamics.In each iteration,oneof these is chosen randomly according to predeﬁned probabili-ties,and()respectively.The above dynamics guarantee the Markovchain designed this way is irreducible(any state is reachablefrom any other state withinﬁnite number of iterations)andaperiodic(the Markov chain does not oscillate in aﬁxed pat-tern)since all of them arestochastic.Figure5:The likelihood based on the number of wronglyclassiﬁed pixels.False positives:in white;true negatives:inshade.NOTE:single-ellipsoid model instead of3-ellipsoidmodel is used for illustration.3.4Multi-human joint likelihood3.4.1Image likelihood by foreground/background sepa-rationSince multiple humans may occlude each other,the imagelikelihood cannot be decomposed into image likelihood asso-ciated with each individual human.Here we use a simple jointlikelihood based on foreground and background separation.(4)Basically,the likelihood reﬂects the number of wronglyclassiﬁed pixels(Fig.5;Equ.4)according to the current stateand the foreground/background of the image frame.Thewrongly classiﬁed pixels()include false positives(,as marked in Fig.5in white region)and true negatives(,as marked in Fig.5in shaded region).False positives re-fer to pixels that do not belong to the foreground but are withinan hypothesized object;true negatives refer to pixels that be-long to foreground but are not within any hypothesized object.One important feature of this likelihood is that it is invariant todepth ordering.By combining Equ.3and Equ.4,Equ.2becomes:(5)is a coefﬁcient which controls the number of correctly clas-siﬁed pixels a human hypothesis has to contribute(excludingthose contributed by others)to increase the posterior probabil-ity.It is mainly decided by the size of human that the applica-tion is interested in.3.4.2Incremental computation of joint likelihoodIn one MCMC iteration only one object may change,sothe computation of the likelihood can be carried out more ef-ﬁciently by incrementally computing it only within the areaassociated with this object and those overlapping with it.We show here how this incremental computation is donewhen creating a new human hypothesis;situation of removinga human or changing the parameter of a human is similar.Fol-lowing the notation in Sec.3.3,is the neighborhood set of anew human.It is obvious that and.where is the number of pixels which satisfy:1)it is in human;2)it belongs to foreground;3)it is not in any human .is the number of pixels which satisfy:1)it is inhuman;2)it belongs to background;3)it is not in any human .4Experiment ResultsWe have tested the approach described above on several data sets and show the results on one outdoor sequence(as the example shown in Fig.1)and one indoor sequence.We set ,,and.In both cases,the Markov chain starts from a null stateand we assume.The results are shown in pairs of inputs and outputs.The inputs are the fore-ground blobs and the head candidates;the outputs are the hu-mans overlaid on the original images with semi-transparent randomly colored masks.Due to the small image size,we have explicitly marked the errors:false alarms in black arrows and miss detections in white arrows.The outdoor sequence is captured from a camera on secondﬂoor.At the far end the humans are very small and the dense edges result in high false alarm rate of head candidates.There are50to100head candidates per frame.We setand5000iterations are used.Result on the example frame in Fig.1is given in Fig.6.Fig.6.(c)shows the log posterior probability of all the iterations.We can see that the posterior climbs very fast and then stays close to optimal solution for ﬁne adjustment.Fig.7shows the results for more frames in the same sequence.The errors usually happen at the far end and some of them are also difﬁcult for a human observer.The indoor sequence has a lower viewpoint(cameramounted on a tripod on the ground)and the inter-occlusion is more severe.The average number of head candidates is around 100.We set and5000iterations are used.Fig.8 shows the results on some of the frames.The people in the scene cause signiﬁcant illumination changes both on the car-pet and on the pillars which causes our background subtrac-tion method to give erroneous foreground blobs.We manually removed those false foreground detections to isolate the dif-ﬁculty of human segmentation.However,the head candidates were obtained from the real foreground.The missed detections either have too few pixels or are almost completely overlapped with other humans.The computation is affected by the complexity of the scene.More iterations are needed for a scene containing more hu-mans and more occlusion.The computation required for each iteration is also proportional to the size of the foreground re-gion and the number of human objects in the state.As an example,a5000-iteration run on the above reporteddataset(a)(b)(c)Figure6:Experiment result of the image in Fig.1.(a)The input;(b)The result.False alarms are marked with dark ar-rows and miss detections are marked with white arrows(same for the followingﬁgures);(c)(solid line)and(dotted line)over iterations.requires about15seconds on a Pentium41.5G Hz PC,with un-optimized C++code.However,we have observed in both sequences,1000iterations are sufﬁcient for the human objects of larger size to be segmented correctly(as can be seen in Fig.6.(c)).5Conclusion and Future WorkWe have presented an approach to segment individual hu-mans in a high-density scene(e.g.,a crowed)acquired from a static camera.This problem is made difﬁcult by objects be-ing inter-occluded.We deﬁne the problem as a model-based segmentation problem where a human shape model is used.A posterior probability is deﬁned based on a human count prior and a foreground/background separation likelihood function. The MAP solution is found using an MCMC framework.Hu-man head candidates are computed from both the foreground boundaries and the intensity image edges to direct the human hypothesizing process of the Markov chain dynamics.Experi-mental results have shown that this approach is highly promis-ing.In this work,knowledge of various aspects including hu-man shape,human height,camera model,and image cues in-cluding human head candidates,foreground/background sepa-ration are integrated in a Bayesian framework.Same method can also be used in other similar problems.5.1Future workThe work described here can be improved in various aspects easily since the MCMC is an open framework.We just list a few important ones here:(a)(b)(c)Figure7:More results on the outdoor sequence.Left column:input;right column:output.Some widely used techniques can be implemented in theMarkov chain dynamics such as having more jump stepsin the early stage and having more diffusion steps in thelater stage;deciding which object to remove according toaﬁtness value,etc.Currently we only use foreground/background separationas likelihood measurement.It will not be effective ifmany people are entirely in the interior of the foreground.Other likelihood cues(e.g.edge,color and motion)maybe helpful.Currently we don’t use any temporal information exceptfor background subtraction.The framework should beextended to make temporal estimation.Human model can be enhanced to handle closer rangecases where the articulation of the limbs cannot be wellcaptured by our current model.However,the low dimen-sionality of the model has to be ensured.References[1]G.Borgefors,Distance Transformations in Digital Images,CVGIP,34,344-371,1986.[2]J.Canny,A Computational Approach to Edge Detection,IEEETrans.on PAMI,V ol.8,No.6,Nov.1986.[3] A.M.Elgammal and L.S.Davis,Probabilistic Framework forSegmenting People under Occlusion,Proc.of Int.Conf.on Com-puter Vision,Vancouver,Canada,2001.(a)(a)(c)Figure8:Results on the indoor sequence.Left column:input;right column:output.[4]S.Haritaoglu,D.Harwood and L.S.Davis,W4:Real-TimeSurveillance of People and Their Activities,IEEE Trans.onPAMI,Vol.22,No.8,2000.[5]R.Hartley and A.Zisserman,Multi View Geometry,pp.191,Cambridge Press,2000.[6]M.Isard and J.MacCormick,BraMBLe:A Bayesian Multiple-Blob Tracker,Proc.of Int.Conf.on Computer Vision,Vancouver,Canada,2001.[7]N.T.Siebel and S.Maybank,Fusion of Multiple Tracking Al-gorithm for Robust People Tracking,Proc.of European Conf.onComputer Vision;LNCS2353,pp.373-387,2002.[8]H.Tao,H.S.Sawhney and R.Kumar,A sampling algorithmfor tracking multiple objects,Proc.of Workshop of Vision Algo-rithms,with ICCV99.[9]L.Tierney,Markov chain concepts related to sampling algo-rithms,Markov Chain Monte Carlo in Practice,Chapman andHall,pp.59-74,1996.[10]Z.W.Tu and S.C.Zhu,Image Segmentation by Data-DrivenMarkov Chain Monte Carlo,IEEE Trans.on PAMI,vol.24,no.5,2002.[11]T.Zhao,R.Nevatia and F.Lv,Segmentation and Tracking ofMultiple Humans in Complex Situations,Proc.of IEEE Conf.onComputer Vision and Pattern Recognition,Kauai,Hawaii,2001.[12]S. C.Zhu,R.Zhang,and Z.W.Tu.Integrating Top-down/Bottom-up for Object Recognition by Data Driven MarkovChain Monte Carlo,Proc.of IEEE Conf.on Computer Vision andPattern Recognition,Hilton Island,North Carolina,2000.。