Long Range Iris Acquisition System for Stationary and Mobile Subjects

Transcription

1 Long Range Iris Acquisition System for Stationary and Mobile Subjects Shreyas Venugopalan 1,2, Unni Prasad 1,2, Khalid Harun 1,2, Kyle Neblett 1,2, Douglas Toomey 3, Joseph Heyman 1,2 and Marios Savvides 1,2 Abstract Most iris based biometric systems require a lot of cooperation from the users so that iris images of acceptable quality may be acquired. Features from these may then be used for recognition purposes. Relatively fewer works in literature address the question of less cooperative iris acquisition systems in order to reduce constraints on users. In this paper, we describe our ongoing work in designing and developing such a system. It is capable of capturing images of the iris up to distances of 8 meters with a resolution of 200 pixels across the diameter. If the resolution requirement is decreased to 150 pixels, then the same system may be used to capture images from up to 12 meters. We have incorporated velocity estimation and focus tracking modules so that images may be acquired from subjects on the move as well. We describe the various components that make up the system, including the lenses used, the imaging sensor, our auto-focus function and velocity estimation module. All the hardware components are Commercial Off The Shelf (COTS) with little or no modifications. We also present preliminary iris acquisition results using our system for both stationary and mobile subjects. 1. Introduction Iris pattern based biometric systems, for surveillance and access control have been gaining a lot of popularity over the recent years. Several such systems are being deployed in highly secure checkpoints across the globe, the primary reason being the uniqueness of iridal patterns across people. Over the years, many works in literature have proposed several feature extraction and comparison schemes for this biometric and they have reported very high recognition rates 1 Department of Electrical and Computer Engineering (ECE), Carnegie Mellon University, Pittsburgh, USA 2 CyLab Biometrics Center, Carnegie Mellon University, Pittsburgh, PA-15213, USA 3 Mauna Kea Infra-Red, Hilo, HI-96720, USA toomey@mkir.com (for e.g. see [15][19]). In all iris pattern based biometric systems, a major concern is the acquisition of very well focused eye images in which relevant iris features are discernible. The acquisition process often requires significant cooperation from the subject whose eye is being imaged. One has to position oneself at a pre-defined location, at a pre-defined distance from the camera and provide sufficient near infra-red (IR) illumination for acquisition. The need for such cooperation is due to the limited capture volume of the system i.e. the volume of space in front of the image acquisition system within which the user s iris is of acceptable quality. Once the iris of the user is within this volume, the user typically remains in that position with limited motion until the system acquires a good quality image. An example of a widely used commercial iris acquisition device is the LG IrisAccess4000 [8]. It uses voice prompts to direct the user into the capture volume. In general, with such systems, the positioning process can seem unintuitive for some users and can result in failure to acquire results. Other systems that have been proposed, involve less cooperation from the users. A good example is the Iris-On-the-Move system (IOM) that was developed by Sarnoff Corporation [18]. Iris patterns are captured while users walk through a portal fitted with near-ir illumination panels. The subject stand-off required by the system is 3 meters. This acquisition system is a fixed focus system with a reported narrow depth of field of 5cm. Compared to traditional desktop/wall mount systems (such as those marketed by Panasonic [10], LG [8] and others), this has the advantage of an increased stand-off distance and reduced level of cooperation from the subject. However, iris acquisition fails if a users iris is not acquired through a fixed, small capture volume. Also, this work proposes a modular approach to increase the height of the capture volume. Multiple cameras are stacked one above the other so that the iris can be captured irrespective of the height of the user. The extra hardware and custom lenses increase the cost of the system. Another category of acquisition systems involves the use of pan-tilt-zoom cameras. These cameras alleviate the constraint of a fixed capture volume. Early attempts at using these cameras are reported by Oki IrisPass-M [9], Sensar R1 [17], Wheeler et al [22]. These systems are based on the use of multiple cam /11/$ IEEE

2 eras - a wide angle scene camera to detect the face/eye in the scene and the second camera with a high magnification lens, specifically aimed to resolve the fine patterns in the iris. Depth estimation is performed using a stereo-camera setup. This information helps in estimating the position of the user in 3D space and hence the focus and pan/tilt settings. Venugopalan and Savvides [21] use a commercial off the shelf (COTS) pan tilt zoom camera to track the faces of subject and to acquire irises when the subject is still. The subjects irises may be acquired from stand-off distances of upto 1.5 meters. They use a single camera setup to acquire both the face and iris from subjects of different heights. The Retica Eagle Eye system [13] uses a scene camera, a face camera, and two iris cameras, which account for its large form factor. The capture volume of this setup is larger compared to the systems described so far, yielding a 3 x 2 x 3m capture volume with increased stand-off (average of 5m). AOptix [7] described a system to perform iris recognition at a stand-off of 2 meters. The system can be used to enroll as well as verify users whose heights vary between 0.9 and 1.9 meters. They use adaptive optics with a multi-stage, realtime closed loop control system, in order to find the subject within a capture volume of depth 1 meter. In this paper we propose a novel long range iris acquisition system that has been designed and developed to acquire pristine quality iris images from stand-off distances of up to 12 meters from stationary or mobile subjects. The range of the system developed will be longer than those proposed in literature so far. In addition, since our system uses Commercial Off The Shelf (COTS) equipment, the cost of the system will be extremely low compared to similar systems that employ several custom made devices. Another advantage of our system is the co-location of the illumination panel and imaging sensor. This facilitates a setup that may be deployed quickly to any location without any worry about the illumination on the subject s side. The pixel resolution of the acquired iris will consistently be greater than 150 pixels across the diameter. The system acquires both the face and the iris using a single camera. In order to track mobile subjects, we have incorporated an area scan camera in addition to the high resolution biometric imager. The function of this camera is to track the subject during his/her motion and keep him/her centered in the frame of the high resolution imager (which captures both the face and the iris with the required resolution). In this way, the iridal patterns of the person may be compared against an existing database and one can determine whether he/she poses a possible security threat. For mobile subjects, our current experiments are being performed at walking speeds of under 0.6m/s. In the following sections of the paper we describe the various components that make up our system. We present results from our acquisition experiments and conclude by outlining current work to enhance the performance of the system. 2. Proposed System In this section we describe briefly the important hardware and software components that make up our system. The hardware modules comprise of the high magnification telephoto lens, the imaging sensor and the auto-focus module. The software components consist of focus estimation, subject speed estimation and integration of focal length adjustment and image acquisition Telephoto Lens and Imaging Sensor First, we briefly describe the sensor and the choice of optics for the system. The requirement of our system is the acquisition of both the face and the iris of the subject of interest with sufficient resolution Telephoto Lens The required focal length of the lens is calculated based on the magnification we need so that, for a given subject standoff distance from the lens, we obtain reasonable resolution across the iris. The typical human iris has a diameter of 12mm [16]. Assuming we need approximately 200 pixels across the iris for the recognition module-as suggested in [12]-the magnification required may be determined as shown below. Typical diamter of the human iris, h o = 12mm Image size required, h i = 200pixels = 200 (side of a sensor pixel)mm = 200 ( )mm = 1.29mm The side of the sensor pixel is obtained from the specifications of the camera we use (see section 2.1.2). Hence the magnification required is at least M = h o = 1.29 = (1) h o 12 Consider f to be the focal length of the lens and D to be the distance of the object (here the subject of interest) from the front of the lens, then the required f is determined as f = M D 1 + M = = 776mm (2) where we have taken D = 8m as the subject stand-off distance from the system. In our system we use the Canon 800mm lens [3] as this would satisfy the above focal length requirement. Additionally, the reader should note that for a value of D greater than this, a higher focal length would be required and would necessitate the use of focal length extenders such as those seen in [2]. However, the use of

3 Figure 1: Variation in focal length required with subject stand-off distance to obtain a 200 pixel resolution across an acquired iris. Figure 2: Variation in iris resolution in pixels with subject stand-off distances, when an 800mm lens is used for imaging. such extenders is known to increase aberrations during image acquisition. Hence, for the purpose of this paper, we perform our imaging experiments without the use of such extenders. Figure 1 shows the variation in focal length with subject stand-off distance to obtain a 200 pixel resolution across an iris, using eqn(2). Figure 2 shows the pixel resolution obtained at various distances using the 800mm lens High Resolution Sensor For the purpose of this paper, we chose a camera with a very high pixel count - the Canon 5D Mark II [4]. A higher pixel count ensures that both the face and iris may be acquired with the required resolution, from a single captured frame. This camera has a full frame sensor with dimensions, 36mm x 24mm. As was shown in the previous section, we use a focal length of 800mm for our imaging requirements. Rearranging the terms in eqn(2), we see that for a value of f = 800mm and stand-off distance of D = 8m, the magnification M = f = 0.11 (3) D f From this value, we can determine the field of view at 8m for the given sensor size to be 32.4cm 21.6cm. However at D = 4m, the field of view is calculated to be 14.4cm 9.6cm. Due to the decrease in the longer dimension of the field of view, at distances closer than 4m it becomes difficult to accommodate the entire face within a frame. Our aim in this paper, as mentioned previously, is to acquire enrollment quality iris images (with pixel resolution 200 pixels) using this device. Due to this constraint as well as constraints on focal length used (section 2.1.1) and the diminishing field of view at shorter stand-off distances, in this paper, we have limited our capture range to 4m 8m. If we relax the pixel resolution required across the iris, then images may be captured at distances of 4m 12m using our current setup (see Figure2). The camera is operated in the portrait mode so that the longer dimension of the field of view corresponds to the length of the face Auto-focus module In order to ensure good focus across all acquired images, we have designed and developed a library of auto-focus routines for use with the system. The Canon Application Programming Interface (API) [5] does not provide easy control over the auto-focus hardware within the camera body. As a result we opted to modify an existing RS232 lens control module manufactured by Birger TM Engineering [1] for the purpose of this system. This ensures we have control over the lens, independent of the camera body. The auto-focus routine developed for the purpose of this paper includes a focus measure function based on spatial gradients. The focus measure over the frame is given by the mean of the magnitude of the two-dimensional gradient at every pixel position within the frame. If f measure represents the focus measure over the frame I of dimension m n, then where, f measure = 1 mn I i,j = (δii,j δx I i,j is the pixel intensity at (i, j). m i=1 j=1 ) 2 + n I i,j (4) ( δii,j 2.3. Speed Estimation for Mobile Subjects δy ) 2 (5) In the case of a mobile subject, we estimate his/her speed as he/she walks towards the system. We assume a constant velocity model for the purpose of this paper. The focus position of the system is set at a checkpoint A at a distance

4 D A from the system, upon initialization. As the subject approaches the system and crosses A, the system starts a timer and moves the focus position to a checkpoint B at a distance D B (see 3). Once the subject crosses this position too, due to the constant velocity assumption, we can estimate the speed and can predict the subjects position C at any point in time. We use this information track the subjects focus with time. Figure 3: Experimentally determined relationship between focus encoder steps of the Birger TM Engineering module we use and the stand-off distance of the subject from the lens. In order to perform focus tracking effectively, the position of the subject has to be mapped onto a stepper motor focus encoder value within the Birger TM mount [1]. Figure 3 shows a graph of distance from system in meters vs. encoder steps, which was experimentally determined. From this graph, we notice that for very small variation in distance, the relationship between distance and focus encoder position can be assumed to be linear. Since we are interested in fine focus adjustment for motion across position C (see Figure 3), we use a linear mapping scheme between distance and encoder steps for our experiments. Once the speed has been estimated, we move the focal length to the location C at which an iris image with the desired resolution may be obtained. Given the resolution required in pixels, the image size h i may be estimated as in eqn(1). Since we are using a lens with focal length f = 800mm, the maximum distance from the system D C at which an iris of this resolution can be acquired, may be obtained from eqns(1) and (2) as, D C = f(h o + h i ) h i (6) Once this is set, we capture a set of images continuously while the subject moves past C. This burst capture mode includes a fine focus tracking procedure as mentioned. Along with the estimated subject speed, this procedure also requires knowledge of the minimum time interval between two consecutive shutter activations of the imaging device. For the Canon 5D Mark II this value is 5 frames per second. In other words, the minimum time interval in the burst mode during which no images can be acquired (between two shutter activations) is 0.2 seconds. Hence, the subject distance from system, D i+1 for each consecutive frame may be determined as, ( ) DB D A D i+1 = D i t B t A i = 1, 2,..., N (7) where D 1 = D C. The linear mapping scheme mentioned earlier enables us to convert this value into an equivalent focus encoder step at each consecutive frame. In addition to the system components that have been detailed above, another essential component is the face tracking component. This is necessary in order to keep the subject within the frame of the Canon 5D Mark II. A wide angle scene camera tracks the subject during the entire acquisition process from location A, all the way till the multiple high resolution images are captured. The tracking is achieved by means of a standard Kalman filter implementation which is found in works such as by [20]. The face tracker directs the motion of a pan/tilt mechanism integrated with the system Infra-Red Illumination Panel It is seen that in most irises, a lot of incident light in the visible spectrum is absorbed and some of it is reflected off the cornea. However, most of the light in the infra-red wavelength, incident on the iris, is reflected back and can be imaged by the sensor. Boyce et al [14] has shown experimentally the validity of this argument and concludes that imaging the iris in the infra-red wavelength provides most of the discriminating information for an iris recognition system. We use a standard infra-red LED source such as the one seen in [6] for our experiments. Four panels of these low power LEDs are optimally aligned such that the individual beams superimpose over the required acquisition range. This provides sufficient illumination to image the necessary discriminating information within the iris texture. The system is also fitted with a filter wheel to switch between the visible spectrum and the infra-red spectrum as required. For the visible block filter, we use an 850nm bandpass filter since, from our experiments we see that most of the iris texture is clearly discernible when imaged around this wavelength. Figure 5 shows the system that was built using the components described. 3. Acquisition Results In this section we present high resolution face and eye images captured using the system described in the previous section. We also compare the eye images captured, to those captured using conventional iris acquisition systems.

5 Figure 4: As the subject approaches our system, and crosses variable checkpoints A and B, we estimate his/her speed. Once this is done, the focus position is set to a position C at which we are assured to obtain an iris of required resolution (eqn. (6)). A number of in-focus images are then acquired by changing the focus continuously based on subject distances estimated using eqn. (7) Iris Acquisition from stationary subjects (a) Our first set of experiments with this system involved a simple focus estimation stage and acquisition of eye images when a subject stands at an arbitrary location in front of the system. During this experiment, the subject is tracked and when he/she is still, the auto-focus module uses the focus measuring scheme outlined in section 2.2 and sets the focus position of the lens to an appropriate value. Figure 6 shows an example of a face that was captured during this experiment and Figure 7 shows a few eye images cropped out from such faces at different distances from the camera. We compare the acquired images with images acquired using standard iris acquisition systems in Figure 9. We see that using the proposed system, one can achieve enrollment quality iris images from a much greater stand-off distance compared to conventional systems Iris Acquisition from mobile subjects (b) Figure 5: The figure shows the system that was designed to capture images of static or mobile subjects from a large stand-off distance. Here we show the images that were acquired when the focus position of the lens was tracked automatically based on estimated subject speed (section 2.3). Figure 8 shows iris images that were captured from a subject moving at speeds of between 0.3m/s and 0.6m/s. From all the experiments conducted, we decided that a sufficient number of images to be acquired as the subject passes through the location C is 10. Of these 10 images, an average of 4 images is always in focus. The remainder of the images has erroneous focus settings due to minor variations in the speed of the subject, and motion blur which is to be expected in a real world setting. The requirement of 10 images is not a drawback, since the entire acquisition process when the subject is passing through C takes less than two seconds. Our cur-

6 (a) Figure 6: An example of a face captured during our experiments at stand-off distances between 4-8 meters. This face was captured from a distance of 7 meters. Once the face is acquired, we detect the eye regions and crop them out. Examples of such cropped out eye regions with iris pattern visible are shown in Figure 7. rent work involves measures to decrease the exposure time during image capture so that motion blur may be minimized and to better tune the focus tracking to the subject s motion. This will help reduce the number of burst images to be acquired. 4. Conclusion In this paper, we ve outlined the design and development of a system to acquire high resolution face and iris images from mobile (or stationary) subjects using a single imaging sensor. For mobile subjects, we have included an optional sensor fitted with a wide angle lens to track his/her movements during the acquisition process. As the subject moves towards the system, a kalman filter based face tracking component moves a pan/tilt mechanism. This is so that the face is always centered in the frame of the high resolution sensor. During the course of the subjects motion, we estimate (b) Figure 7: (a) shows an iris image capture from a subject standing still at a distance of 6 meters from our system and (b) shows an image from the same subject at 7 meters. his/her speed using a real time (lower resolution) feed from the high resolution sensor. Using this estimated value, the focus of the system is finely tuned to the subjects motion and a set of high resolution images are captured when he/she passes through a desired location. This desired location (location C in Figure 3) is determined based on the pixel resolution required across the acquired iris (see Figure 2). A sample of images acquired from both stationary and mobile subjects can be seen in Figure 7 and Figure 8. Our current work involves developing a library of iris segmentation and feature extraction functions for use with this system. By imposing a 200 pixel requirement on iris images for enrollment, the subject can be at a maximum stand-off of 8 meters (Figure 2). However, we feel that the pixel requirement can be relaxed during the testing stage during which lower resolution iris images may be acquired and compared with the enrollment images. In that case, the-

7 References (a) (b) Figure 8: (a) and (b) show images from two different capture sessions as the subject walked at speeds between 0.3 and 0.6 m/s. The images were captured at a distance of approximately 6m. oretically, as can be seen in Figure 2, test images may be acquired for distances even beyond 10 meters. An optimum set of filters will have to be designed to pick out, consistently, the most discriminating features from both these resolutions. Also, we are in the process of tuning the focus estimation module, so that a lesser number of burst images need be acquired to obtain a clear iris from the subject. Eventually we hope to use this system in a completely unconstrained environment to enroll and recognize subjects on the move from a large stand-off distance. 5. Acknowledgement We thank the Biometrics Identity Management Agency (BIMA) for supporting this research through ARL grant W911NF [1] Birger engineering. last accessed on 8/11/2011. [2] Canon 1.4X Focal Length extender available at. consumer/eos_slr_camera_systems/lenses/ extender_ef_1_4x_ii last accessed on 8/11/2011. [3] Canon 800mm f5.6 USM telephoto lens available at. products/cameras/ef_lens_lineup/ef_ 800mm_f_5_6l_is_usm last accessed on 8/11/2011. [4] Canon EOS 5D Mark II available at. usa.canon.com/cusa/consumer/products/ cameras/slr_cameras/eos_5d_mark_ii last accessed on 8/11/2011. [5] Canon SDK available at. com/cusa/consumer/standard_display/sdk_ homepage last accessed on 8/11/2011. [6] Infra-Red LED 850nm available at. sparkfun.com/products/9469 last accessed on 8/11/2011. [7] Insight Iris Recognition System available at. last accessed on 8/11/2011. [8] Lg IrisAccess available at. ps/products/irisaccess4000.html. [9] OKI Irispass available at. iris/. [10] Panasonic BM-ET330 available at. http: // bm-et300_demo/index.html last accessed on 8/11/2011. [11] PIER-T iris Acquisition System available at. last accessed on 8/11/2011. [12] A. I Iris Image Interchange Format. [13] F. Bashir, P. Casaverde, D. Usher, and M. Friedman. Eagleeyes: A system for iris recognition at a distance. In Technologies for Homeland Security, 2008 IEEE Conference on, pages , may [14] C. Boyce, A. Ross, M. Monaco, L. Hornak, and X. Li. Multispectral iris analysis: A preliminary study51. In Computer Vision and Pattern Recognition Workshop, CVPRW 06. Conference on, page 51, june [15] J. Daugman. Probing the uniqueness and randomness of iriscodes: Results from 200 billion iris pair comparisons. Proceedings of the IEEE, 94(11): , nov [16] J. Forrester, A. Dick, P. Mcmenamin, and W. Lee. The Eye: Basic Sciences in Practice. W.B. Saunder, London, [17] G. Guo, M. Jones, and B. P. A System for Automatic Iris Capturing. Mitsubishi Electric Research Laboratory http: // [18] J. Matey, O. Naroditsky, K. Hanna, R. Kolczynski, D. LoIacono, S. Mangru, M. Tinker, T. Zappia, and W. Zhao. Iris on the move: Acquisition of images for iris recognition in less constrained environments. Proceedings of the IEEE, 94(11): , nov

8 (a) (b) (c) (d) Figure 9: This figure compares iris images using (a) our system from 9m, with images captured using the (b) LG system [8] from 30cm, (c) the PIER [11] from 15cm and (d) the iris on the move system by Sarnoff [18] from 3m. [19] P. Phillips, K. Bowyer, P. Flynn, X. Liu, and W. Scruggs. The iris challenge evaluation In Biometrics: Theory, Applications and Systems, BTAS nd IEEE International Conference on, pages 1 8, oct [20] P. Turaga, G. Singh, and P. Bora. Face tracking using kalman filter with dynamic noise statistics. In TENCON IEEE Region 10 Conference, volume A, pages Vol. 1, nov [21] S. Venugopalan and M. Savvides. Unconstrained iris acquisition and recognition using cots ptz camera. EURASIP J. Adv. Signal Process, 2010:38:1 38:20, February [22] F. Wheeler, A. Perera, G. Abramovich, B. Yu, and P. Tu. Stand-off iris recognition system. In Biometrics: Theory, Applications and Systems, BTAS nd IEEE International Conference on, pages 1 7, oct