Personal Authentication Based on Hands Natural Layout Miguel Adán, Antonio Adán, Roberto Torres, Andrés S. Vázquez, Gloria Bueno Escuela Superior de Informática. UCLM. Spain {Miguel.Adan, Antonio.Adan, Roberto.Torres, Andres.Vazquez, Gloria.Bueno}@uclm.es Abstract This work is addressed to develop a hands biometric system for verification and recognition goals. The method is based on three keys. First we use a hand’s natural layout as intrinsic properties of each individual. Consequently, neither hand-pose training nor a pre-fixed position is required in the registration process. Secondly, we define a set of features without using typical image processing (i.e. segmentation, filtering, etc) because these are defined on a part of the hand’s contour. Thirdly, instead of common methods that register one hand, we use left and right hands in our approach. As a consequence of this, ratios FAR and FRR improves meaningfully. The paper shows the experimentation and results of the method for more than 4200 real samples taken in a secondary school. The results are good enough to consider this biometric system for future security/control applications. 1 Hand-based biometric Unfortunately, security issues are becoming more and more of increasing interest in today's society. The identification of individuals based on biometrics is an important component towards this goal. Besides, most biometric systems depend on passwords and codes one of the most reliable methods. This is due to the fact that they are based on each individual's inherent characteristics. Currently, there is a large number of biometric systems making use of face recognition, voice analysis, iris pattern, fingerprints, hand geometry, etc. Hand-based biometric technologies are getting very popular for control purposes, such as access to buildings, airports, nuclear plants, and Olympic stadiums. They are suitable for massive use because of their low processing time and real time response. Unlike biometric systems based on fingerprints recognition and iris pattern, users are not reluctant to hand biometric ones. Most features related to a human hand are relatively invariant and peculiar (although, not unique) to each individual. That is why these systems have been commercially used more for verification than for identity recognition. These systems make use of only one hand, usually the right one, from which the features are obtained by different methods, such as hand geometry, [1], [2], palm-prints [3], [4], [5], finger crease [6] and deformable model, [7]. Moreover, the image acquisition is usually accomplished in controlled environments. These environments are formed by a platform composed of a set of pegs, which enable the hand position to be fixed, and a mirror to obtain the up side of the hand [1], [8], [9], [10]. The main drawbacks of these systems are: a) the required upkeep due to damage, spoiling and dirtiness of the platform and mirrors and b) the required training and co-operation of the individuals to place the hand in the position constrained by the pegs. These systems are improved by those based on 3D reconstructions [11] albeit of using expensive and complicated sensor systems. Our project explores, in a novel way, the analysis of both hands geometry data. As far as we are concerned, systems that make use of right and left hands do not exist. Furthermore, the hand pattern is related to the implicit axis within the natural hand layouts, which allows us to design a simple and affordable system free of pegs and easy to use where the user does not have to be trained. The only requirement is to place the hands with outstretched fingers. The following sections present: features extraction based on the natural hand layout (section 2), similarity measure (section 3), verification results (section 4), conclusions and further work (section 5). 2 Pattern definition in a Natural Reference System In most hand biometric systems the hand is placed at a pre-fixed position [1], [9], [12]. It can be said that the hand is positioned with respect to a universal reference system (i.e. image reference system). Contrary to this idea, the hand features that we define are relative to their own reference axes called Natural Reference System (NRS). This reference system is based on two invariant points of the extended hand: the positions of the middle and thumb ends (see Figure 1 a). Let Y X O , , be the image coordinate system, O being the centre of the image and suppose an image containing an extended hand D. Let 1, 2, 3, 4, 5 be the labels of pinkie, ring finger, middle finger, index finger and thumb respectively. The Natural Reference System of D is defined after the following steps: I) Find the straight line , r r fitted to the skeleton of 3. II) Set ' , ' , ' Y X O with O’=O, Y’ being parallel to r and passing by O’ and X’ being normal to Y’ by O’. III) Find 0 0 , y x P , D y x x Min x ) ' , ' ( : ' 0 on 5 in the system ' , ' , ' Y X O . IV) Move ' , ' , ' Y X O to 0 , 0 ' ' y O obtaining the Natural Reference System ' ' , ' ' , ' ' Y X O . Natural placing of the fingers on a completely extended hand is theoretically invariant and personal. Figure 1 b) illustrates three different samples for a hand, where the relative placing of the fingers remains invariant. Note the particular position for both index and MVA2005 IAPR Conference on Machine VIsion Applications, May 16-18, 2005 Tsukuba Science City, Japan 3-36 160