Robust Object Tracking under Appearance Change Conditions

Transcription

1 Inernaional Journal of Auomaion and Compuing 7(1), February 2010, DOI: /s Robus Objec Tracking under Appearance Change Condiions Qi-Cong Wang Yuan-Hao Gong Chen-Hui Yang Cui-Hua Li Deparmen of Compuer Science, Xiamen Universiy, Xiamen , PRC Absrac: We propose a robus visual racking framework based on paricle filer o deal wih he objec appearance changes due o varying illuminaion, pose varianions, and occlusions. We mainly improve he observaion model and re-sampling process in a paricle filer. We use on-line updaing appearance model, affine ransformaion, and M-esimaion o consruc an adapive observaion model. On-line updaing appearance model can adap o he changes of illuminaion parially. Affine ransformaion-based similariy measuremen is inroduced o ackle pose varianions, and M-esimaion is used o handle he occluded objec in compuing observaion likelihood. To ake advanage of he mos recen observaion and produce a subopimal Gaussian proposal disribuion, we incorporae Kalman filer ino a paricle filer o enhance he performance of he resampling process. To esimae he poserior probabiliy densiy properly wih lower compuaional complexiy, we only employ a single Kalman filer o propagae Gaussian disribuion. Experimenal resuls have demonsraed he effeciveness and robusness of he proposed algorihm by racking visual objecs in he recorded video sequences. Keywords: Visual racking, paricle filer, observaion model, Kalman filer, expecaion-maximizaion (EM) algorihm 1 Inroducion Wih he developmen of compuer science, arificial inelligence, elecronic echnology, ec., he inelligen visual surveillance promises a wide range of applicaions, such as naional defense, human-compuer ineracion, securiy, and space exploraion. Therefore, more and more aenion is paid o he researches of visual racking [1 3]. I is a key echnology of he inelligen visual surveillance and is a very imporan research opic in compuer vision. Tracking objec whose appearance is changing due o varying illuminaion, pose varianions, scale changes, and occlusions is one of he core challenges in visual racking. Since Bayesianheory-based paricle filer has been proven o be successful for non-linear and non-gaussian esimaion problems and effecive for solving racking problems [3 5], i has been exensively sudied for visual racking, in which he observaion model and he imporance re-sampling are wo main componens. The observaion likelihood impacs is robusness, and he imporance re-sampling has grea effecs on is efficiency. In order o deal wih he changes in he environmen, some racking algorihms [6 8] employ he appearance model based on color disribuion o compue he observaion likelihood. The racked arge is modeled using color disribuion from he firs frame and kep unchanged during racking. Is disadvanage is unable o rack over long ime. In [9], he weighed average of he curren and new arge color disribuion was employed o updae he appearance model. Insead of using color disribuion, Gaussian mixure models have been used o model he racked arge agains he changing appearance in many racking algorihms. In [10], hisogram-based rackers based on Gaussian disribuion were proposed, bu i mus deermine he number of componens in advance. In [11], appearance sub- Manuscrip received March 12, 2009; revised June 1, 2009 This work was suppored by Naional Naural Science Foundaion of China (No ) and he 985 Innovaion Projec on Informaion Technique of Xiamen Universiy ( ). space model learned wih principal componen analysis can furher enhance he robusness of visual racking under variaions in pose and lighing condiions. However, his model is objec specific and mus be rained in advance [12]. In [12], an online appearance model ha was a 3-componen mixure was proposed for a robus visual racking. Boh he appearance model and he racking parameers were derived by expecaion-maximizaion (EM) algorihm. In [13, 14], he 3-componen mixure was furher improved by employing hree appearance models ogeher, in which proper appearance models for he curren racking siuaion could be auomaically seleced. However, boh of hem only used a fixed number of componens. In [15], an on-line densiybased appearance model for visual racking was proposed, in which all of he parameers of mixure Gaussians could be auomaically compued by a mean-shif algorihm and represen he densiy funcion of he racked region very accuraely wih minimal componens. I is more flexible han he radiional Gaussian mixure models. In a convenional paricle filer, sysem ransiion is used as he proposal disribuion. However, he imporance resampling does no ake ino accoun he mos recen observaion from he curren frame, and he dynamic one is very difficul o be accuraely modeled in he image sequences. This ofen leads o poor priors for he resampling sep. I is no he opimal proposal disribuion. Generally, many paricles are required o esimae he poserior probabiliy densiy properly. In [16 18], he principle of he Kalman filer was inroduced ino he filering process of paricle filer, where paricle sae was updaed by using he mos recen observaions. As a resul, fewer paricles were required o represen he disribuion properly for racking he objec conours. In his paper, we presen a robus objecs racking algorihm based on he improved paricle filer, in which he adapive observaion model and Kalman filer are used. The adapive observaion model employs on-line appear-

2 32 Inernaional Journal of Auomaion and Compuing 7(1), February 2010 ance model whose parameers are updaed by an online EM algorihm, affine ransformaion, and M-esimaion o adap o he changed appearance of he racked objec due o unsable lighing condiion, pose variaions, scale changes, view-poin changes, and occlusions. In order o ake advanage of he mos recen observaions, a Kalman filer is inegraed ino he paricle filer for designing a subopimal Gaussian proposal disribuion. Experimenal resuls on he indoor and oudoor video sequences demonsrae he effeciveness and robusness of he proposed algorihm. The res of his paper is organized as follows. Secion 2 inroduces he basic algorihm of he radiional paricle filer. In Secion 3, we presen he main componens of he adapive observaion model in deail. Secion 4 describes how o embed Kalman filer in a paricle filer. Some resuls of experimens on objec racking are shown in Secion 5. Finally, conclusions are briefly drawn in Secion 6. Paricle filer provides a robus racking framework, as i models uncerainy. Paricle filer wih a re-sampling process is called sampling imporance re-sampling (SIR) paricle filer. The SIR algorihm is composed of hree main seps: sampling, weighing, and resampling, as shown in Fig Tradiional paricle filer Paricle filer solves non-linear and non-gaussian sae esimaion problems in Mone Carlo simulaion using imporance sampling, in which he poserior densiy is approximaed by he relaive densiy of paricles in a neighbourhood of sae space. Paricle filer deals wih he racking problem based on he moion model and on he observaion model x = f(x 1, ε ) (1) Fig. 1 The sampling imporance re-sampling paricle filer y = h(x, v ) (2) where ε and v are only supposed o be independen whie noises. y 0: is defined as he hisory sequence of he random variables. Our problem consiss in compuing he condiional densiy p(x y 0:) of he sae x a ime, which can be esimaed hrough predicion and updae recursively. We realize predicion according o he following equaion: p(x y 0: 1) = p(x x 1)p(x 1 y 0: 1)dx. (3) The observaion y enables us o updae his predicion using he Bayes rule p(x y 0:) = p(y x )p(x y 0: 1). (4) p(y x)p(x y0: 1)dx The sample se {(x n,, q n,) n=1,2,,n }, where x is he paricle sae, and q is he weigh associaed o he paricle, is used o esimae he densiy p(x y 0:) by he formulaion p(x y 0:) N q n,δ(x x n,) (5) n=1 where δ is he dela funcion. The weigh q is calculaed as p(y x n,)p(x n, x n, 1) q n, q n, 1 (6) s(x n, x n, 1, y ) where s is he imporance densiy funcion. In he sampling sep, new paricles are generaed by drawing from he imporance proposal disribuion. In he weighing sep, he weighs associaed wih paricles are evaluaed by means of he observaion model. Then, he weighs are normalized so ha he weighs add up o uniy. In he re-sampling sep, new paricles are drawn from he disribuion represened by he previous paricle se, and all weighs associaed wih he paricles are se o be equal o weigh 1/N. The re-sampling sep is crucial in he implemenaion of paricle filering because wihou i, he variance of he paricle weighs increases quickly. Because re-sampling is uniform and sysem ransiion is used as he proposal disribuion, he radiional paricle filer is difficul o mainain muliple modes of paricle se afer several updaes and can resul in he sample impoverishmen problem. Therefore, a large number of paricles are required o obain high accuracy. The major problems in designing paricle-filer-based visual racking algorihms include eliminaing paricle impoverishmen and designing reasonable and efficien observaion model and moion model. Generally speaking, obaining an accurae moion model is difficul by using only images. Thus, people are more ineresed in he researches of he observaion model and he impoverishmen problem. In fac, many facors can affec he effecive observaion of he racked arges in images, such as varying illuminaion, pose varianions, scale changes, occlusions, and so on. The observaion model is required o deal wih hose facors adapively so ha he objec can be racked sably and persisenly.

3 Q. C. Wang e al. / Robus Objec Tracking under Appearance Change Condiions 33 3 Adapive observaion model 3.1 EM-algorihm-based mode seeking Mode seeking is a very imporan mehod in feaurespace-based analysis of images. In compuer vision, significan feaures correspond o he regions of high-densiy values and he regions of low-densiy values are of no ineres for he feaure space analysis [19]. A local mode ha is he locaion of a local maximum in a probabiliy densiy funcion is used o describe his ype of he region of high-densiy values. Feaure space analysis wih probabiliy densiy is o seek he local modes of he densiy. In general, here are several local maxima in a probabiliy densiy funcion simulaneously, and hey are close o one anoher. Therefore, unreliable mode deecion is easily yielded in he sense and more likely o happen a low signal o noise raios. In [20], Gaussian-kernel-based mean-shif was proved o be an EM algorihm. Mean-shif, which always poins oward he direcion of maximum increase in he probabiliy densiy funcion, is a robus mode seeking approach. The magniude of mean-shif vecor is small near local maxima, which correspond o he regions of high-densiy values. Moreover, we ge more precise analysis. On he conrary, he magniude of mean-shif vecor is large in he regions of low-densiy values. Mean-shif-based mode seeking is o locae he zeros of he normalized probabiliy densiy gradien wihou acually compuing his densiy. This procedure is similar o an adapive gradien ascen mehod. Compared wih he radiional gradien-based mehods, he advanages of he mean-shif are an adapive procedure whose sep size does no need o adjus and he guaraneed convergence. Therefore, we use Gaussian-kernel-based mean-shif, i.e., EM algorihm, o seek he local modes in he probabiliy densiy of he racked objec appearance. 3.2 On-line updaing appearance model In paricle-filer-based visual racking, how o model he appearance of he objec is paricularly criical o is adapion o varying illuminaion. Therefore, our adapive observaion model uses he on-line updaing appearance model ha is able o updae auomaically and is similar o he on-line densiy-based appearance model [15]. I is assumed ha a ime, he pixels in he image region of he racked objec are independen of each oher, and he iniial probabiliy densiy is a Gaussian mixure model having n modes. We allocae a Gaussian disribuion for each mode. Assume ha a ime + 1, a new mode is observed as he (n + 1)- h mode. Wih he inegraion of his mode and a learning rae α, he probabiliy disribuion a ime +1 is compued by he formulaion ˆf +1(x) = 1 α (2π) d 2 α n (2π) d 2 σ n K i exp( 1 σ i D2 (x, µ i, σ))+ i exp( 1 2 D2 (x, µ n +1, σ n +1 )) where D 2 (x, µ i, σ i ) = (x µ i ) T (σ i ) 1 (x µ i ) is he Mahalanobis disance i = 1, 2,, n and K i, µ i, σ i are weigh, mean, and covariance of Gaussian disribuion (7) N(K i, µ i, σ i ), respecively. Le n N(K, i µ i, σ) i = 1 k i exp( 1 (2π) d 2 σ i D2 (x, µ i, σ)). i (8) If ω m is a corresponding weigh of a sample belonging o he i-h mode, he probabiliy densiy disribuion of his mode is wrien as f(µ i, σ i ) = M ω mn(x m, µ i, σ). i (9) m=1 We would like o find he mean µ i and he covariance σ, i which maximize he value of he above equaion. This can be done ieraively using EM algorihm [11]. From he Jensen s inequaliy, we ge log f(µ i, σ) i G(µ i, σ, i q 1, q 2,, q N ) = N ln( ωmn(xm, µi, σ) i (10) ) qa q a a=1 where q a is a consan and N a=1 qa = 1, qa 0. We denoe he curren esimaed values of he mean µ i and he covariance σ i as µ i(l). To esimae he mean, he EM algorihm is repeaed as follows unil convergence: E sep. Firs, keeping µ i(l) and σ i(l) fixed, we would like and σ i(l) o find {q a} o maximize G. From he above equaion, we ge ha wih equaliy sign, he maximum is achieved for q a = ωmn(xm, µi(l) M m=1, σ i(l) ) ω mn(x m, µ i(l), σ i(l) ). (11) M sep. To maximize G wih respec o µ i and σ i, we can use he following fixed-poin ieraion scheme for fixed {q a}. We would like o minimize G ha depends on he mean µ i and he covariance σ i : From g(µi, σ i ) µ i g(µ i, σ i ) = µ i(l+1) = M q a ln N(x m, µ i, σ). i (12) a=1 = 0, we ge M m=1 ω mn(x m, µ i(l), σ i(l) ω mn(x m, µ i(l), σ i(l) ) )x m. (13) To find he mean of he new mode in ˆf +1(x), we perform EM algorihm ieraions unil he convergence for each. The above ieraion process for he mean esimae is equivalen o he mean-shif updae equaion for he Gaussian kernels. In order o updae he covariance, we employ Hessian H = ( T ) ˆf +1 o esimae covariance [15]. The updaed probabiliy densiy funcion a ime + 1 is hen wrien as ˆf +1(x) = 1 (2π) d 2 n +1 k i σ+1 i exp( D2 (x, µ i +1, σ+1)). i (14)

4 34 Inernaional Journal of Auomaion and Compuing 7(1), February Pose variaions and scale changes handling The observaion model is designed based on he on-line updaing appearance model. Assume ha he number of pixels in he appearance model is d. The image observaion of he racked objec a ime can be represened as Z = {Z (1), Z (2),, Z (d)} (15) where Z(j) is he observed gray value in he image a he pixel poin j. We can denoe he appearance model a ime 1 by M k = {k i, µ i, σ i }. The observaion likelihood funcion over he sae x is wrien as p(z x ) = d n ( k(j)n(z i (j), µ i (j), σ(j))) i (16) j=1 where N(Z (j), µ i (j), σ(j)) i is Gaussian disribuion. We would like o generalize he applicaion of he on-line updaing appearance model for visual racking under pose variaions and scale changes by inroducing four parameers of affine ransformaion. Le ϕ denoe he affine ransformaion funcion and denoe A by {a x, b y, r, θ }, where a x, b y, r, and θ correspond o x, y ranslaions, scale, and roaion angle a ime, respecively. Then, he racked objec a ime can be deermined by he following affine ransformaion: ( ) ( ) ( ) ( ) a T cos θ sin θ a a x b T = r + sin θ cos θ b b x ( ) (17) a where X = is he pixel poin in he racked objec b ( ) a ime, and X = a b is he pixel poin afer affine ransformaion. Then, he observaion of he racked objec can be compued as Z = I(ϕ(X, A )) (18) where I(ϕ(X, A )) is he pixel gray value observed in he image a he pixel poin ϕ(x, A). Afer obaining he observaion corresponding o he sae x, he observaion likelihood is represened as p(z A ) = d p(z (i) A ) = d n ( k(j)n(z i (j), µ i (j), σ(j))). i (19) j=1 3.4 Occlusion handling In general, if he pixels wihin he racked objec are occluded, hey have large image differences, and can be reaed as ouliers. In order o rack he occluded objec, we also adop he M-esimaion echnique [14], which is defined as 1 ρ(γ (j)) = 2 (η(j))2, if η (j) c c η (j) 1 (20) 2 c2, oherwise where η (j) = [Z (j) µ i 1(j)]/σ i 1(j), i = 1,, n 1 is a mach error, and c is a hreshold ha conrols he oulier rae. If he condiion η (j) c is no saisfied, he pixel j is viewed as an oulier. 4 Kalman filer embedded in paricle filer Similar o Kalman paricle filer [16 18], we also employ Kalman filer o improve paricle filer, in which a subopimal Gaussian proposal disribuion is consruced and propagaed hrough Kalman filer. I is differen from he Kalman paricle filer ha he Kalman observaion updae only uses he mean and covariance of paricle saes. Therefore, he compuaional complexiy is reduced grealy. In paricle filer, he paricle sae ransiion from 1 o can be expressed wih he dynamical model x = F x 1 + Gε (21) where F is referred o as he sae ransiion marix, G is he sysem noise marix and ε is a noise erm. This noise erm is a Gaussian disribuion [ ] wih zero mean and a covariance marix Q = E ε ε T. The observaion z can be expressed in erms of he sae wih he observaion model z = Lx + γ (22) where L is he covariance marix referred o as measuremen noise covariance marix, and γ is he noise [ of he ] observaion wih a covariance marix R = E γ γ T. The radiional paricle filers do no employ his ype of observaion model. We would like o inroduce i ino paricle filer hrough Karman filer. The filering process has wo sages. One is he updae sage, and he oher is he predicion sage. Updaed paricle sae is expressed by Is covariance marix is defined as ˆx = E[x z,, z 1]. (23) ˆP = E[(x ˆx )(x ˆx ) T ]. (24) Prediced paricle sae is denoed by x = E[x z,, z 1]. (25) The covariance marix associaed wih his paricle sae is defined as follows: P = E[(x x )(x x ) T ]. (26) The i-h updaed and prediced paricles which have a sae, (i) a covariance marix and a weigh can be described by ˆP = {ˆx (i) (i), ˆP, ŵ (i) (i) } and P = { x (i) (i), P, w (i) }, respecively. We ge he observaion vecor z = [a, b, r, θ ] hrough he weighed average of he prediced paricle sae. z = N w (i) L x (i). (27) In general, i is assumed ha he arge is in uniform moion for visual racking. So, he i-h paricle sae a ime is

5 Q. C. Wang e al. / Robus Objec Tracking under Appearance Change Condiions 35 x (i) = [a (i), b (i), a (i) 1, b(i) 1, r(i), θ (i) ]. We assume ha he objec is moving according o he following equaions: a +1 = a + (a a 1) + ε a (28) b +1 = b + (b b 1) + ε b (29) r +1 = r + ε r (30) θ +1 = θ + ε θ. (31) Then, we ge he sae ransiion marix F, he sysem noise marix G, and he covariance marix L easily F = G = L = The noise vecors are defined as follows: ε (i) γ (i) = [ε (i) a, ε (i) b, ε(i) r, ε (i) θ ]T (32) = [γ (i) a, γ (i) b, γ(i) r, γ (i) θ ]T. (33) The covariance marices associaed wih ε (i) defined as Q (i) = σ (i)2 ε I and R (i) defined as he uni marix and σ (i) γ Kalman gain is compued by K (i) = (i) P L T (L = σ (i)2 γ and γ (i) are I, respecively. I is w (i). Therefore, he (i) P L T + R (i) ) 1. (34) The Kalman observaion updae process can be wrien as ˆx (i) = x (i) ˆP (i) ŵ (i) + K (i) (z L x (i) ) (35) = (I K (i) (i) L) P (36) p(z (i) ˆx (i) ). (37) The Kalman predicion process is given by where ˆx (i) x (i) +1 = F ˆx (i) + Gε (i) (38) P (i) (i) +1 = F ˆP F T + GQG T (39) w (i) +1 p(z(i) x (i) +1 ) (40) is drawn from he updaed paricle se ˆx (i). The final paricle sae of he racked objec a ime is he weighed average of updaed paricles using x ou = N ŵ (i) ˆx (i). (41) 5 Experimenal resuls Our racker was implemened in visual C++ environmen and esed on a wide variey of real-word video sequences. The es videos condiions included illuminaion changes, facial pose variaions, and parial occlusions. The proposed algorihm ook approximaely 15 fps on PIV 2.8 G CPU worksaion wihou opimizing he code and was applied o images of sizes and , respecively. The racking resul is shown wih a whie bounding box. The racker was iniialized manually by placing a recangle region in he firs image. In he following secions, we illusrae some represenaive video sequences. 5.1 Tracking a man s body under illuminaion changes The video sequences shown in Figs. 2 and 3 are a man moving in an oudoor environmen, in which he appearance of he objecs has changes and is of low conras during racking. These condiions make racking difficul, so he observaion model of he racker mus be able o adap o hose condiions for robus racking. Fig. 2 shows he resuls of racking a man s body under low illuminaion changes. The resuls show ha he racker can rack he man s body robusly. In Fig. 3, he racking resuls for he man under illuminaion changes are shown. In his sequence, he appearance of he arge varies frequenly because of changes in illuminaion. The resuls have shown ha he racker adaps o his well. Fig. 2 (a) Frame # 1 (b) Frame # 88 (c) Frame # 100 (d) Frame # 164 (e) Frame # 188 (f) Frame # 239 Some racking resuls under low illuminaion condiions

6 36 Inernaional Journal of Auomaion and Compuing 7(1), February 2010 (a) Frame # 7 (b) Frame # 100 (a) Frame # 2 (b) Frame # 59 (c) Frame # 115 (d) Frame # 162 (c) Frame # 239 (d) Frame # 456 (e) Frame # 188 (f) Frame # 194 (e) Frame # 580 (f) Frame # 681 Fig. 3 ions Some racking resuls under varying illuminaion condi- 5.2 Tracking a man0 s face under pose variaions and scale changes Fig. 4 Some racking resuls under pose varianions and scale changes The gray-scale image sequence shown in Fig. 4 is a man changing his head poses in an office environmen, in which he objec involves scale changes and pose variaions. The appearance of he racked objec changes over ime. The racking resuls show ha our racking algorihm can succeed in racking hroughou he sequence under pose variaions and scale changes. The main reason is ha we have inroduced affine ransformaion ino he adapive observaion model. 5.3 (a) Frame # 1 (b) Frame # 9 (c) Frame # 17 (d) Frame # 18 (e) Frame # 25 (f) Frame # 29 Tracking a car under parial occlusions The video sequences shown in Figs. 5 and 6 are a car parially occluded by rees. I can be seen in Fig. 5 ha he racker can succeed in racking he car afer a period of occlusions. The main reason for robus racking under parial occlusions is ha he racker incorporaes M-esimaion ino he adapive observaion model. In he las sequence, he appearance of he objec changes frequenly because of occlusions of he rees. Fig. 6 shows ha he occlusion handling can succeed in adaping o he appearance changes when parial occlusions happen. However, he racker can lose he arge finally because of he heavy occlusions. Fig. 5 Some racking resuls under parial occlusion condiions

7 Q. C. Wang e al. / Robus Objec Tracking under Appearance Change Condiions 37 [3] M. Isard, A. Blake. Condiional densiy propagaion for visual racking. Inernaional Journal of Compuer Vision, vol. 29, no. 1, pp. 5 28, (a) Frame # 2 (b) Frame # 18 [4] M. S. Arulampalam, S. Maskell, N. Gordon, T. Clapp. A uorial on paricle filers for online nonlinear/non-gaussian Bayesian racking. IEEE Transacions on Signal Processing, vol. 50, no. 2, pp , [5] P. Yang, W. Wu, M. Moniri, C. C. Chibelushi. A sensorbased SLAM algorihm for camera racking in virual sudio. Inernaional Journal of Auomaion and Compuing, vol. 5, no. 2, pp , [6] D. Comaniciu, V. Ramesh, P. Meer. Kernel-based objec racking. IEEE Transacions on Paern Analysis and Machine Inelligence, vol. 25, no. 5, pp , (c) Frame # 31 (d) Frame # 74 [7] P. Pérez, C. Hue, J. Vermaak, M. Gangne. Color-based probabilisic racking. In Proceedings of European Conference on Compuer Vision, Lecure Noes in Compuer Science, Springer, London, UK, vol. 2350, pp , [8] Q. Wang, J. Liu, Z. Wu. Objec racking using geneic evoluion based kernel paricle filer. Lecure Noes in Compuer Science, Springer, vol. 4040, pp , Fig. 6 (e) Frame # 98 (f) Frame # 104 Some racking resuls under heavy occlusion condiions 6 Conclusions We have proposed a mehod for visual racking in video using he improved paricle filer based on adapive observaion model and Kalman filer. The adapive observaion model employs he on-line updaing appearance model, affine ransformaion, and M-esimaion echnique o handle illuminaion changes, pose variaion, and parial occlusions, respecively. The adapive observaion model was incorporaed in he paricle filer, in which Kalman filer was adoped o produce he subopimal re-sampling proposal disribuion. Experimenal resuls on real-world video sequences have demonsraed ha he proposed algorihm is robus o rack objecs in some complex environmens. Acknowledgemen The firs auhor hank Prof. Ji-Lin Liu, Eryong Wu, and Ye-Hu Shen in Zhejiang Universiy for he help and encouragemen. References [1] N. Wang, G. Y. Wang. Shape descripor wih morphology mehod for color-based racking. Inernaional Journal of Auomaion and Compuing, vol. 4, no. 1, pp , [2] D. Xu, M. Tan, X. Zhao, Z. Tu. Seam racking and visual conrol for roboic arc welding based on srucured ligh sereovision. Inernaional Journal of Auomaion and Compuing, vol. 1, no. 1, pp , [9] K. Nummiaro, E. Koller-Meier, L. Van Gool. An adapive color-based paricle filer. Image and Vision Compuing, vol. 21, no. 1, pp , [10] S. J. McKenna, Y. Raja, S. Gong. Tracking colour objecs using adapive mixure models. Image and Vision Compuing, vol. 17, no. 3 4, pp , [11] M. J. Black, A. D. Jepson. EigenTracking: Robus maching and racking of ariculaed objecs using a view-based represenaion. Inernaional Journal of Compuer Vision, vol. 26, no. 1, pp , [12] A. D. Jepson, D. J. Flee, T. F. El-Maraghi. Robus online appearance models for visual racking. IEEE Transacions on Paern Analysis and Machine Inelligence, vol. 25, no. 10, pp , [13] A. Li, Z. Jing, S. Hu. Robus observaion model for visual racking in paricle filer. Inernaional Journal of Elecronics and Communicaions, vol. 61, no. 3, pp , [14] S. Hu, G. Liang, Z. Jing. Robus objec racking algorihm in naural environmens. In Proceedings of he 2nd Inernaional Conference on Naural Compuaion, Springer, Xi an, PRC, vol. 4222, pp , [15] B. Han, L. Davis. On-line densiy-based appearance modeling for objec racking. In Proceedings of IEEE Inernaional Conference on Compuer Vision, IEEE, Beijing, PRC, vol. 2, pp , [16] P. Li, T. Zhang, A. E. C. Pece. Visual conour racking based on paricle filers. Image and Vision Compuing, vol. 21, no. 1, pp , 2003.

8 38 Inernaional Journal of Auomaion and Compuing 7(1), February 2010 [17] C. Shen, M. J. Brooks, A. Van Den Hengel. Augmened paricle filering for efficien visual racking. In Proceedings of IEEE Inernaional Conference on Image Processing, IEEE, vol. 3, pp , [18] Y. Saoh, T. Okaani, K. Deguchi. A color-based racking by Kalman paricle filer. In Proceedings of he 17h Inernaional Conference on Paern Recogniion, IEEE, vol. 3, pp , [19] D. Comaniciu, P. Meer. Mean shif: A robus approach oward feaure space analysis. IEEE Transacions on Paern Analysis and Machine Inelligence, vol. 24, no. 5, pp , [20] M. A. Carreira-Perpinan. Gaussian mean shif is an EM algorihm. IEEE Transacions on Paern Analysis and Machine Inelligence, vol. 29, no. 5, pp , Qi-Cong Wang graduaed from Nanjing Universiy of Aeronauics and Asronauics, PRC in He received he M. Sc. degree from Zhejiang Universiy of Technology, PRC in 2004, and he Ph. D. degree from Zhejiang Universiy, PRC in He is currenly an assisan professor in he Deparmen of Compuer Science a Xiamen Universiy, PRC. His research ineress include compuer vision, image processing, informaion fusion, and arificial inelligence. qcwang@xmu.edu.cn (Corresponding auhor) Yuan-Hao Gong received he bachelor degree from Tsinghua Universiy, PRC in He is currenly a maser suden a Deparmen of Compuer Science, Xiamen Universiy, PRC. He is a member of IEEE and ACM. His research ineress include image processing and undersanding, compuer vision, and paern recogniion. yuanhaogong@acm.org Chen-Hui Yang received he Ph. D. degree in Zhejiang Universiy, PRC. He is currenly a professor in he Deparmen of Compuer Science a Xiamen Universiy, PRC. His research ineress include compuer visions, image processing and analysis, arificial inelligence in raffic and is applicaion. chyang@xmu.edu.cn Cui-Hua Li received he Ph. D. degree in he Insiue of Arificial Inelligence and Roboics from Xi an Jiaoong Universiy, PRC. He is currenly a professor in he Deparmen of Compuer Science a Xiamen Universiy, PRC. His research ineress include compuer visions, image processing and analysis, wavele ransformaion heory and is applicaion. cuihuali@xmu.edu.cn