Products & Services

A nature based fusion scheme for multimodal biometric person identity verification at a distance

Description
A nature based fusion scheme for multimodal biometric person identity verification at a distance
Published
of 4
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
  A nature based fusion scheme for multimodal biometric person identity verificationat a distance Ho Chiung Ching , C.Eswaran Faculty of Information Technology,Multimedia UniversityCyberjaya, Selangor, Malaysia{ccho,eswaran}@mmu.edu.my  Abstract   — This paper presents a multimodal biometricverification scheme for face, gait and speech data as inspiredby how verification is done at a distance in the natural world.  Keywords-nature base, multimodal biometric verification, fusion I.   I  NTRODUCTION Person identity verification is important for securingaccess to scarce or restricted resources. As security threatsincreases, the need to ascertain the claim of identity that a person makes becomes increasingly important. Personidentity verification is the process of proving an identityclaim (“I am person X”) and can be done through variousapproaches. Traditional approaches calls for the person toeither have knowledge that proofs his identity, e.g. passwords or possesses an identity token which proofs whohe is; for example a pre-verified building access card. The problem for both approaches is that passwords can beforgotten, while an identity token can be misplaced or stolen.A combination of both approaches , for example a buildingaccess card which requires a passkey ; can be more secure but nonetheless is still vulnerable to theft or a lapse inmemory. As such, biometrics is a better solution. Biometrics[1] is the usage of unique physical or physiological attributesfrom a person as proof of their identity. Physical biometricmodalities include facial features, speech, thumbprint, palm print and even the shape of a person’s ears and lips.Behavioral biometric modalities can be found in a person’sgait and cadence, as well as the person’s signature. All of these biometric features are unique to a person, and aremostly time invariant. Recent development in biometric based person identity verification has progressed from asingle modality to a mesh of multiple biometric modalities.A multimodal biometric authentication scheme is morerobust than a single modal authentication scene as there isgreater tolerance for signal loss or degradation as there ismore than one biometric modality considered [1]. In thisrespect, loss of integrity for single modal biometrics throughsignal loss or degradation is overcome by considering theother remaining modality. A composite result is achievedthrough fusion at either feature level or decision level for themultimodal biometric modalities.Person identity verification happens in the natural worldas well as electronically – the age old challenge for a password from a distant stranger before permission isgranted to proceed is conceptually identical to manycomputer based identification system, ranging fromIdentification Friend or Foe (IFF) systems on military radar systems to video based automatic person verification system.In nature, verification often takes place based on audio andvisual features. For example, chimpanzees can tell mother and son pairings based on visual clue as shown in Parr andde Waal’s work [2]. A person can also classify another  person as being familiar (known) or unfamiliar based onclues from the way the other person walks (gait), the way the person talks (speech) and finally a decision is made based onhow the person looks like (facial). The accuracy of theclassification increases as the physical distance between two person decreases – this happens as the facial features becomemore prominent, and at the same time the speech features becomes more distinct.II.   P ASS L ITERATURE    A. Human Visual System inspired video-based verification Widespread usage of computer vision for security relatedapplications has encouraged deep interest in modeling thehuman visual system (HVS). One particular area of interestwould be in how aspects HVS can be employed to improvevideo surveillance, especially in how the HVS is able totrack shape, movement and color in spite of a wide range of illumination. Peerasathein [3] modeled the ventral streamwhich contains the primary visual cortex using a neuralnetwork classifier to enable video based object classification.Kim [4] represented the HVS as a parameterized monte carlomarkov model in his work to classify 3D objects. The HVS’s physical and behavioral was modeled by Carnec and Barbain the context of facial recognition, which yielded goodresults. Current HVS based approaches for authenticationeither approximate the physical structures within the HVS or the behavioral aspect of HSV.Person verification is a task that is markedly differentfrom person identification. Identification calls for a probe to be examined against a gallery sample in a database, whereas 2009 International Conference on Signal Acquisition and Processing 978-0-7695-3594-4/09 $25.00 © 2009 IEEEDOI 10.1109/ICSAP.2009.2894  verification is where a probe is presented and the task it todetermine if the probe claim to identity is accurate.Facial features, gait and speech are a natural fit to videoas these biometric modalities are easily captured on video.As such, combinations of the three aforementionedmodalities are often used for person identity verification.  B. Face-based identity verification Face based identity verification is arguably the mostmature of biometric modalities. Facial features are invariantfrom person to person, and are easily extracted from either still images or a stream of images taken from a video feed.Face based verification is one of the most mature modality,as shown in the huge amount of literature available on it.Face verification algorithm are closely related to facerecognition algorithms, and includes techniques such asEigenfaces and Fisherfaces[5], Gabor classifiers [6],Bayesian [7] and many others. Of late, there has been effortsto develop algorithmns that are meant for verification per say, instead of being modified from face recognitionalgorithm. Yan’s work in 2007 [8] introduces the concept of a balanced threshold between pre-learned global thresholdand optimal class specific threshold. C. Gait-based identity verification Gait is the way in which a person walks. Nixon reportedin 2003 [9] that features from a person’s gait in uniqueacross individuals. It is observable from a distance, and isdistinct at a range that surpasses other biometric modalitiessuch as face or voice. Gait verification is a classification problem – subjects are classified as clients (intra class) or impostors (inter class).Verification methods are looselygrouped into two groups – either silhouette based or model based. Silhouette based techniques makes use of thesubject’s silhouette which is applied with techniques such assymmetry [10] and body shape [11] analysis. Model basedtechniques incorporate shape and dynamic features from gaitto enable verification. These techniques include motion [12]and body part trajectories [13].  D. Speech- based identity verification Speech verification is the act of verifying a person’sidentity through their speech. Speech has latent propertieswhich are invariant and unique from person to person.Speech verification can be divided into categories – textdependent, in which the user has prior knowledge of the textspoken, and text independent, where the user is free to sayany phrase to verify their identity. Popular techniques for speech verification include adapted Gaussian Mix models[14], Hidden Markov Model [15] and wavelets.  E. Multimodal biometric fusion Multimodal biometrics seeks to improve on therobustness of single biometric systems. It can be done in avariety of ways; such as capturing several different biometricfeatures from several different biometric sources; usingdifferent procedures to yield different information, differentalgorithms can be applied for different matching scores, andconsolidating several rules for decision making. Dahelsummarized this in his paper [16], which is presented in thefollowing diagram.Figure 1: Levels of Multimodal biometric fusionInformation fusion is a method in which the prediction of various sources of information is consolidated in one presentation. Figure I summarized the different levels of fusion available. Decisions during fusion can be made indifferent ways, such as weighted, equal or dynamic. Conradis of the opinion that the weighted option is the most flexiblefor multimodal biometric fusion [17].  F. Face and gait biometrics Facial features and gait are both unobtrusive biometricmodalities. A use often times need not know that they areunder observation, especially if the video cameras areconcealed. Facial biometrics and gait biometrics further complement each other in their effective range, gait workswell when observed from a distance while facial biometricsdepend on the quality of the image of the face; short rangeclose-ups of the face will give excellent results. Chellapa[18] and Kale [19] demonstrated that this combination isviable and gives excellent results. G. Face and speech biometrics Face and speech multimodal biometrics are commonelements in a video based person verification system. Wheneither the audio or video stream is compromised, the other modality is present as a backup. Certain features from faceand speech; for example lip shape and spoken words arenaturally closely coupled. These features are used as inputfor a classifier such as a GMM. Kittler and Poh has ablyshown that the combination of face and speech brings aboutgood results in a natural way. Biometricsensor ISensor    Fusion   Biometricsensor IIBiometricsensor IFeature   Fusion   Biometricsensor IIBiometricsensor IDecision Fusion Matcher IBiometricsensor IOpinion   Fusion   Decision I 95  III.   V ERIFICATION FROM A DISTANCE IN THE NATURALWORLD  Verification in the natural world often takes place at adistance. The HVS has an effective range of 500 feet andanything seen beyond this range will lose focus and appear  blurred. At extreme ranges, face based verification is at agreat disadvantage. Lighting also plays a great part at adistance. Extreme backlight will effectively destroy anydistinct facial features and only leave the person’s bodyoutline. At this point of time, verification is better donethrough gait, rather than by face. As distance is reduced;verification is done through speech. Speech carries better through distance as compared to facial features. Finally, at arange of 150 feet or so, a person’s face would be visibleenough so as to allow verification to be done based on the person’s face. This transition in effectiveness of gait, speechand face features for verification is shown in the followingfigure.Figure 2: Progress of biometric modality processing inthe natural worldIV.    NATURE BASED VERIFICATION SCHEME  The nature based authentication scheme for personidentity verification from a distance is inspired by howverification is done in the natural world. At a great distance,gait is the most dominant biometric modality as facialfeatures are obscured and speech is attenuated. At mid point,speech is given a greater weight as speech becomes clearer as the distance between the person who are speaking becomes lesser. At close distance face features are the mostdominant as the HVS relates face to identity.This scheme is a variation from the normal multimodal biometric matching score fusion scheme. The challenge isnot so to find an optimal balance of weights across the 3modalities, but to do so dynamically. Hui describes a similar approach in which he applied fuzzy logic [20] todynamically alter the weights of face fingerprint and speechto simulate variances while capturing these modalities.Cheng also employs adapted weights in Adaboost in [21].Figure 3: Nature based scheme for verification at a distanceusing adaptive weightV.   F UTURE WORK AND CONCLUSION  This paper presents a dynamically adapted weightedmatching score fusion for face, gait and speech data that isinspired by how verification is done in the natural world.Adaptive weights are suggested to simulate the changes infeature perception over a physical distance, as it is done behaviorally by the HVS.R  EFERENCES   [1]   A.K. Jain, A. Ross, and S. Prabhakar, "An Introduction toBiometricRecognition," IEEE Trans. on Circuits and Systems for VideoTechnology, 14:4--20, 2004.[2]   Parr,L.A. and de Waal F.B.M.,”Visual Kin Recognition inChimpanzees”,Nature 399:647, 1999.[3]   Peerasathein, T., Myung Woo and Gaborski, R.S., “BiologicallyInspired Object Categorization in Cluttered Scenes,”Proc. 36 th IEEEWorkshop on Applied Image Pattern Recognition (AIPR2007), 10-12Oct. 2007, pp.117 – 122, DOI 10.1109/AIPR.2007.13[4]   S. Kim, G-J Jang, W.L Lee, and I.S. Kweon, ”How human visualsystems recognize objects - a novel computational model,”Proc. 17 th  International Conference on Pattern Recognition (ICPR2004), 23-26Aug. 2004 pp:61 - 64 DOI 10.1109/ICPR.2004.1334469[5]   P. Belhumeur, J. Hespanha, and D. Kriegman, “Eigenfaces vs. Fisher-faces: Recognition using class speci fi c linear projection,” IEEE Trans.Pattern Anal. Mach. Intell., vol. 19, no. 7, pp. 711–720, Jul. 1997[6]   C. Liu and H. Wechsler, “Gabor feature based classi fi cation using the FaceFacefeatureSpeechGait featureextractor Speechfeatureextractor GaitMatchingscoreMatchingscoreMatchingscoreAdaptive weightMatching score fusion 96  enhanced Fisher linear discriminant model for face recognition,”IEEE Trans. Image Process., vol. 11, no. 4, pp. 467–476, Apr. 2002.[7]   B. Moghaddam, T. Jebara, and A. Pentland, “Bayesian face recogni-tion,” Pattern Recognition., vol. 33, pp. 1771–1782, 2000.[8]   S. Yan, X.Dong and X.Tang, “Face Veri fi cation With BalancedThresholds”, IEEE Trans. on Image Processing,16:262-268,2007.[9]   M. S. Nixon, J. N. Carter, M. G. Grant, L. Gordon, and J. B. Hayfron-Acquah, "Automatic recognition by gait", Sensor Review, vol. 23,200 pp. 323-331.[10]   J. B. Hayfron-Acquah, M. S. Nixon, and J. N. Carter, "Automatic gaitrecognition by symmetry analysis", Pattern Recognition Letters, vol.24, 2003, pp. 2175-2183.[11]   R. T. Collins, R. Gross, and J. Shi, "Silhouette-based humanidentification from body shape and gait", Proc. 5th IEEE Int'l Conf.Automatic Face Gesture Recognition, Washington, DC, USA,2002, pp. 366-71[12]   H. Ning, L. Wang, W. Hu, and T. Tan, "Articulated model based people tracking using motion models", Proc. 4th IEEE Int'l Conf.Multimodal Interfaces, Pittsburgh, PA, USA, 2002, pp. 383-8.[13]   D. Meyer, J. Posl, and H. Niemann, "Gait classification with HMMsfor trajectories of body parts extracted by mixture densities", Proc.British Machine Vision Conf. (BMVC 98), Southampton, UK, 1998, pp. 459-68[14]   D. Reynolds. T. Quatieri and R. Dum, "Speaker Verification Us-ing Adapted Gaussian Mixture Models". Digital Signal Processing,VOL. 10, NO. 1-3,2000,pp. 19-41.[15]   J. MariClhoz, D. Genoud. F. Bimbo1 and C. Mokbel, "Client/WorldModel Synchronous Alignement for Speaker Verification".Proceedings of EUROSPEECH. Budapest. 1999, pp. 1979.1982(Vol. 5).[16]   Dahel, S.K., and Xiao, Q., “Accuracy performance analysis of multimodal biometrics” Information Assurance Workshop, 2003.IEEE Systems, Man and Cybernetics Society , 18-20 June 2003 pp.170-173[17]   C. Sanderson, Information fusion and person verification usingspeech and face information. September 2002[18]   R.Chellapa,A.K. Roy-Chowdurry, and A.Kale,“Human Identi fi cationusing Gait and Face”, Proc. IEEE Conference on Computer Visionand Pattern Recognition (CVPR2007), Dec. 2007, pp. 1-2, doi:10.1109/CVPR.2007.383523.[19]   A. Kale, A. Roy-Chowdhury, and R. Chellappa, “Towards a viewinvariant gait recognition algorithm”, Proc. IEEE AVSS, pp. 143– 150, 2003.[20]   Hui, H.P.-S.,Meng, H.M.,and Man-Wai Mak, “Adaptive WeightEstimation in Multi-Biometric Verification using Fuzzy LogicDecision Fusion”, Proc. 2007 IEEE International Conference onAcoustics, Speech and Signal Processing, 2007, pp. I-501 - I-504,doi: 10.1109/ICASSP.2007.366726.[21]   L. Cheng, J.P. Zhang,J. Yang, and J.Ma , “An Improved AdaBoostAlgorithm Based on Adaptive Weight Adjusting”, Proc. 3 rd  International Conference on Advanced Data Mining andApplications, 2007, pp. 625-632, doi: 10.1007/978-3-540-73871-8_60. 97
Search
Tags
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks