Liveness detection method and apparatus of video image

This application claims the priority of Korean Patent Application No. 10-2007-0015530, filed on Feb. 14, 2007, in the Korean Intellectual Property Office, the disclosure of which is incorporated herein in its entirety by reference.

1. Field of the Invention

The present invention relates to a liveness detection method and apparatus, and more particularly, to a liveness detection method and apparatus which detect a real object by distinguishing a two-dimensional (2D) flat object from a three-dimensional (3D) non-flat object in a video image.

2. Description of the Related Art

Recently, many studies on biometrics have been conducted. Biometrics is widely applied to biometric authentication systems using face or iris recognition, mobile phone security, mobile commerce, physical/logical access control, e-commerce, and the like. However, no attention is paid to liveness detection.

Biometric authentication systems are vulnerable to forgery since fingerprints, irises, and faces can be easily forged. Hence, overcoming vulnerability to forgery is becoming an important issue for biometric system security. The kind of forgery described above is called ‘spoofing,’ which is carried out using a series of processes to intrude into a biometric authentication system using fake biometric samples.

Liveness detection apparatus resistant to spoofing is essential to tackle such spoofing. In particular, it is required to implement liveness detection technology using a non-intrusive method and without additional hardware.

Conventional real face detection methods are classified into a method using movements and three-dimensional (3D) technology, a user interaction method, and a method using additional hardware.

The method using movements and 3D uses a structure from motion method (SFM) in order to estimate the 3D-depth of a tracked object. This depth information is used to distinguish a real facial image from a photographed image.

The user interaction method tracks facial features and interprets a feature trajectory or a change in form. However, this method is inaccurate since the trajectory of a real facial characteristic image is similar to that of a photographed image. In addition, user manipulation is required to track features of a face, thereby causing inconvenience.

The invention disclosed in U.S. Patent No. 2005/0129286 relates to determining whether a real face has been made by examining the state of the eyes and changes of eye position. However, since conventional imaging hardware takes a photograph of an image 25 or 30 frames per second, it cannot not easily detect eye blinks. Therefore, users have to carefully change focus to meet system requirements.

The inventions disclosed in U.S. Patent Nos. 2002/0183624 and 2005/0134427 are related to the method of using additional hardware. Specifically, the invention disclosed in U.S. Patent No. 2002/0183624 uses an active-type light-emitting device in order to distinguish real samples from fake ones. While this invention is a robust algorithm and does not require user manipulation, it requires special light-emitting hardware that limits its application range. Similarly, the invention disclosed in U.S. Patent No. 2005/0134427 also requires special hardware to measure body temperature.

The present invention provides a liveness detection method and apparatus which can perform strong liveness estimation by distinguishing a two-dimensional (2D) flat object from a three-dimensional (3D) non-flat object using video sequence images and without requiring additional hardware or user manipulation.

According to an aspect of the present invention, there is provided a liveness detection method including detecting an object feature point from a predetermined frame of a video image and initializing a tracking operation; tracking object feature points of subsequent frames of the video image based on the detected object feature point; affine-transforming the detected object feature point and the tracked object feature point of an (n−1)-th frame for an n-th frame and calculating statistical distances between the affine-transformed object feature points and the tracked object feature point of the n-th frame; and calculating a liveness value based on the calculated statistical distances and determining whether the video image is a real object based on the calculated liveness value.

According to another aspect of the present invention, there is provide a liveness detection method including detecting an object feature point from a video image and initializing a tracking operation; tracking object feature points of frames subsequent to a frame from which the object feature point was detected; and affine-transforming the detected object feature point and the tracked object feature point of an (n−1)-th frame for an n-th frame and determining whether the video image is a real object using characteristics of the affine-transformed object feature points and the tracked feature point of the n-th frame.

According to another aspect of the present invention, there is provided a liveness detection apparatus including an initialization unit detecting an object feature point from a predetermined frame of a video image and initializing a tracking operation; a tracking unit tracking object feature points of subsequent frames of the video image based on the detected object feature point; a calculation unit affine-transforming the detected object feature point and the tracked object feature point of an (n−1)-th frame for an n-th frame and calculating statistical distances between the affine-transformed object feature points and the tracked object feature point of the n-th frame; and a detection unit calculating a liveness value based on the calculated statistical distances and determining whether the video image is a real object based on the calculated liveness value.

According to another aspect of the present invention, there is provided a recording medium on which a program for executing the method in a computer is recorded.

The above and other features and advantages of the present invention will become more apparent by describing in detail exemplary embodiments thereof with reference to the attached drawings in which:

FIG. 1 is a schematic block diagram of a liveness detection apparatus according to an embodiment of the present invention;

FIG. 2 is a detailed block diagram of the liveness detection apparatus illustrated in FIG. 1 ;

FIG. 3 is a diagram for illustrating the concept of statistical distances (D 1 and D 2 ) according to another embodiment of the present invention;

FIGS. 4 through 6 compare statistical distances D 1 , D 2 and V 1 of a real object with those of a photographed object according to another embodiment of the present invention;

FIG. 7 is a flowchart illustrating a liveness detection method according to another embodiment of the present invention;

FIG. 8 is a flowchart illustrating a liveness detection method according to another embodiment of the present invention; and

FIG. 9 is a schematic block diagram of an access control system using a liveness detection apparatus, according to another embodiment of the present invention.

The present invention will now be described more fully with reference to the accompanying drawings, in which exemplary embodiments of the invention are shown. The invention may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth therein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the invention to those skilled in the art.

FIG. 1 is a schematic block diagram of a liveness detection apparatus according to an embodiment of the present invention.

Referring to FIG. 1 , the liveness detection apparatus includes an initialisation unit 100 , a tracking unit 110 , a calculation unit 120 , and a detection unit 130 .

The initialisation unit 100 detects an object feature point from a specified frame of an input video image and initialises tracking. In this case, an object and the object feature point may be detected in a first frame or its subsequent frame. This detection process is repeated until the object and its feature point are detected in the video image.

If the object and its feature point are detected in the first frame or its subsequent frame by the initialisation unit 100 , the tracking unit 110 tracks an object feature point of each subsequent frame based on the detected object feature point using a predetermined tracking algorithm such as an algorithm based on an optical flow which is calculated using a Lucas-Kanade-Tomasi (LKT) algorithm or a mean-shift algorithm. A detailed description of the tracking algorithm can be found in “Pyramidal Implementation of the Lucas Kanade Feature Tracker,” J. Y. Bouguet, OpenCV Documentation, Microprocessor Research Labs, Intel Corp., 2000 and “Real-Time Tracking of Non-Rigid Objects Using Mean Shift,” D. Comaniciu, CVPR 2000.

The calculation unit 120 performs affine transformation of each object feature point detected in the first frame or its subsequent frame, and an object feature point of an (n−1) th frame, which is tracked based on the feature point detected in the first frame or its subsequent frame, to an n th frame, i.e., a current frame.

The calculation unit 120 calculates a distance D 1 between the affine-transformed object feature point, which is detected in an initial frame, i.e., the first frame or its subsequent frame, and calculates a variance V 1 of the distance D 1 . In addition, the calculation unit 120 calculates a distance D 2 between the affine-transformed object feature point of the (n−1) th frame, which is a previous frame of the current frame n th frame and an object feature point of the n th frame, i.e., the current frame. The affine transformation and statistical distances D 1 and D 2 and V 1 will be described in detail later with reference to FIG. 3 .

The detection unit 130 calculates liveness values using values of the calculated distances D 1 and D 2 and variance V 1 and determines whether the input video image is a real object or a photographed object based on the calculated liveness values.

FIG. 2 is a detailed block diagram of the liveness detection apparatus illustrated in FIG. 1 .

Referring to FIG. 2 , the liveness detection apparatus includes the initialisation unit 100 , which includes an object detection unit 200 and a feature detection unit 210 ; the tracking unit 110 ; the calculation unit 120 , which includes a feature point normalization unit 220 , an affine transformation unit 230 and a distance calculation unit 240 ; and the detection unit 120 which includes a liveness value calculation unit 250 and a liveness detection unit 260 .

The object detection unit 200 detects an object from a specified frame of an input video image. In an embodiment of the present invention, an object includes a face, and an adaboosting algorithm suggested by J. Vola is used for face detection. A detailed description of face detection is disclosed in “Robust Real-Time Object Detection,” Paula Viola, vol. 57, p. 137-154, no. 2, 2004.

The feature detection unit 210 detects feature points in the object detected by the object detection unit 200 . In the present embodiment, the object detection unit 200 detects object feature points in an (n−1) th frame of the input video image. In particular, the object detection unit 200 detects a facial image in an image of a first frame and selects facial feature points, which reflect major features of a face, in the detected facial image. In this case, the object is a person's face. Examples of the detected feature points include two points respectively corresponding to central positions of the eyes, a point corresponding to the position of a nose tip, two points respectively corresponding to both ends of the lips, and a point corresponding to the uppermost point of the upper lip.

Methods of detecting object feature points include an active appearance model (MM) and an active shape model (ASM). The ASM detects the position of the eyes and determines the positions of feature points using shape information of parts other than the eyes, such as the nose, lips and eyebrows, according to a generic model based on the detected position of the eyes. The MM detects feature points, which can efficiently describe facial features, by taking into consideration texture information in addition to shape information.

In the present embodiment, corners may be detected in order to select better feature points. That is, corners of a detected facial region are detected and feature points are selected accordingly in order to reduce errors which may occur during the detection of feature points.

Corner detection is performed using Equations (1) and (2) below. That is, a matrix G and an eigen value λ m of the matrix G are calculated for each pixel of an image I.

G = ∑ x = p x - w x p x + w x ⁢ ∑ y = p y - w y p y - w y ⁢ [ I x 2 ⁡ ( x , y ) I x ⁡ ( x , y ) ⁢ I y ⁡ ( x , y ) I x ⁡ ( x , y ) ⁢ I y ⁡ ( x , y ) I y 2 ⁡ ( x , y ) ] , ( 1 ) I x = I ⁡ ( x + 1 , y ) - I ⁡ ( x - 1 , y ) 2 ⁢ ⁢ I y = I ⁡ ( x , y + 1 ) - I ⁡ ( x , y - 1 ) 2 , ( 2 )

where I(x, y) indicates a grey image, and (w x ,w y ) indicates the width and height of an image region.

The matrix G and the unique value λ m of the matrix G are calculated, and feature points are selected. The maximum eigen value among the eigen values λ m for each pixel of the image is λ max , and pixels satisfying λ m >0.2λ max are searched for a maximum pixel. In this case, the eigen values λ m of the selected pixels are greater than those of other pixels. In addition, the distance between the selected pixels is maintained to be less than a predetermined value.

A detailed description of corner detection is disclosed in “Good Features to Track,” Jianbo Shi, Proc. IEEE Comput. Soc. Conf. Comput. Vision and Pattern Recogn., p. 593-600, 1994.

The tracking unit 110 tracks feature points of subsequent frames based on the feature points of the object detected by the feature detection unit 210 .

In an embodiment of the present invention, if facial feature points are detected in the first frame by the initialisation unit 100 , the tracking unit 110 calculates an optical flow which minimizes the difference between pixel values of the first frame and a second frame. For example, the tracking unit 110 may obtain a grey level of a template image of a predetermined size based on each of the facial feature points detected in the first frame and a grey level of a template image of the second frame, which corresponds to the template image of the first frame and calculate the optical flow using the difference between respective grey levels of pixels included in the template images. In the present embodiment, the optical flow is calculated using an LKT algorithm. A value of an optical flow v, which minimizes a value of ε(v), can be calculated using Equations (3) through (7) below.

ɛ ⁡ ( v ) = ɛ ⁡ ( v x , v y ) = ∑ x = u x - ω x u x + ω x ⁢ ∑ y = u y - ω y u y + ω y ⁢ ( I ⁡ ( x , y ) - J ⁡ ( x + v x , y + v y ) ) 2 , ( 3 ) I x ⁡ ( x , y ) = I ⁡ ( x + 1 , y ) - I ⁡ ( x - 1 , y ) 2 ⁢ , I y ⁡ ( x , y ) = I ⁡ ( x , y + 1 ) - I ⁡ ( x , y - 1 ) 2 , ( 4 ) G ⁡ ( u x , u y ) = ∑ x = u x - ω x u x + ω x ⁢ ∑ y = u y - ω y u y + ω y ⁢ [ I x 2 I x ⁢ I y I x ⁢ I y I y 2 ] ,

⁢ δ ⁢ ⁢ I ⁡ ( x , y ) = I ⁡ ( x , y ) - J ⁡ ( x , y ) , ( 5 ) b _ = ∑ x = u x - ω x u x + ω x ⁢ ∑ y = u y - ω y u y + ω y ⁢ [ δ ⁢ ⁢ II x δ ⁢ ⁢ II y ] , ( 6 ) v = G - 1 ⁢ b _ , ( 7 )

where I indicates a previous ((n−1) th ) frame, J indicates a current (n th ) frame, u=(u x , u y ) indicates a feature point of the previous frame, v=(v x , v y ) indicates an optical flow, (w x , w y ) indicates a template size of an image for calculating the optical flow, and I x and I y indicate differential image values of I.

The feature point normalization unit 220 normalizes the feature points detected by the feature point detection unit 210 , that is, the feature points of the initial frame, the feature frames of the current frame (n th frame), and the feature frames of the previous frame ((n−1) th frame).

The detected feature points are normalized using Equation (8).

P i ⁡ ( n ) = P i ⁡ ( n ) - P _ ⁡ ( n ) σ P ⁡ ( n ) ⁢

⁢ { P _ ⁡ ( n ) = 1 N ⁢ ∑ i = 1 N ⁢ P i ⁡ ( n ) σ P ⁡ ( n ) = ( 1 N ⁢ ∑ i = 1 N ⁢ P i ⁡ ( n ) · P i ⁡ ( n ) - P _ ⁡ ( n ) · P _ ⁡ ( n ) ) , ( 8 )

where P i (n), n=1, 2, . . . N.

The affine transformation unit 230 performs an affine transformation on the feature points normalized by the feature point normalization unit 220 . In other words, the affine transformation unit 230 performs the affine transformation on the feature points of the initial frame for the current frame (n th frame) and performs the affine transformation on the feature points of the previous frame ((n−1) th frame) for the current frame (n th frame).

S x = 1 n ⁢ ∑ i = 1 n ⁢ x i S y = 1 n ⁢ ∑ i = 1 n ⁢ y i S x ′ = 1 n ⁢ ∑ i = 1 n ⁢ x i ′ S y ′ = 1 n ⁢ ∑ i = 1 n ⁢ y i ′ S xx = 1 n ⁢ ∑ i = 1 n ⁢ x i 2 S yy = 1 n ⁢ ∑ i = 1 n ⁢ y i ⁢ y i S xy = 1 n ⁢ ∑ i = 1 n ⁢ x i ⁢ y i S xx ′ = 1 n ⁢ ∑ i = 1 n ⁢ x i ⁢ x i ′ S yy ′ = 1 n ⁢ ∑ i = 1 n ⁢ y i ⁢ y i ′ S xy ′ = 1 n ⁢ ∑ i = 1 n ⁢ x i ⁢ y i ′ S yx ′ = 1 n ⁢ ∑ i = 1 n ⁢ y i ⁢ x i ′ , ( 9 )

where (x i , y i ) indicates coordinates of feature points determined in the previous frame ((n−1) th frame), and (x i ′, y i ′) indicates coordinates of feature points determined in the current frame (n th frame).

An affine transformation parameter T is determined by six elements (t x , t y , a, b, c and d) as shown in Equation (10).

The biggest difference between a photographed facial image and a real facial image is that the photographed facial image is a 2D flat object without depth changes while the real facial image is a 3D object having relatively large depth changes.

In an ideal imaging condition, ideal orthography, which is a condition without image noise and distortion, and an ideal tracking algorithm without tracking errors and bias, the distance between affine-transformed tracking points different frames is a flat object, that is, zero for a photographed face. On the other hand, this distance is greater for a real facial object than a photographed facial object.

However, the distance in a photographed face is definitely not zero. Therefore, stronger liveness detection must be defined to exclude external factors, such as tracking effects and imaging conditions, and to reflect internal factors such as the difference in 3D depths of a photographed face and a real face.

The affine transformation will now be described with reference to FIG. 3 . First of all, it is assumed that a feature point i of a first frame is P i ( 1 ), where i=1, 2, N (N>3).

First through n th frames are illustrated in FIG. 3 . In addition, the feature point P i ( 1 ) ( 300 ) of the first frame, a feature point P i (n−1) ( 310 ) of the (n−1) th frame, and a feature point P i (n) ( 320 ) of the n th frame are illustrated.

For N>3, the feature point P i ( 1 ) ( 300 ) of the first frame is affine-transformed for a current frame (the n th frame). The affine-transformed feature point of the first frame is indicated by {tilde over (P)} i ( 1 ) ( 305 ).

The feature point P i (n−1) ( 310 ) of a previous frame (the (n−1) th frame) is affine-transformed for the current frame (the n th frame). The affine-transformed feature point of the (n−1) th frame is indicated by {tilde over (P)} i (n−1) ( 315 ).

Referring back to FIG. 2 , the distance calculation unit 240 calculates a distance D i 1 (n) between the affine-transformed feature point {tilde over (P)} i ( 1 ) ( 305 ) of the first frame and the feature point P i (n) of the current frame using Equation (12) below.

D i 1 ⁡ ( n ) =  P i ⁡ ( n ) - P ~ i ⁡ ( 1 )  .

⁢ D 1 ⁡ ( n ) ⁢ ⁢ is ⁢ ⁢ given ⁢ ⁢ by ⁢ ⁢ Equation ⁢ ⁢ ( 13 ) . ( 12 ) D 1 ⁡ ( n ) = 1 K ⁢ ∑ i = 1 K ⁢ D i 1 ⁡ ( n ) , ( 13 )

where D 1 (n) is an average of D i 1 (n), and K is the number of feature points.

D 1 is zero for flat objects. A real value of D 1 (n) is mainly related to an accumulated tracking error. The real value of D 1 (n) increases when a tracked point is located a large distance from its initial position.

In addition, the distance calculation unit 240 calculates a distance D i 2 (n) between the affine-transformed feature point {tilde over (P)} i (n−1) ( 315 ) of the (n−1) th frame and the feature point P i (n) of the current frame using Equation (14) below.

D i 2 ⁡ ( n ) =  P i ⁡ ( n ) - P ~ i ⁡ ( n - 1 )  .

⁢ D 2 ⁡ ( n ) ⁢ ⁢ is ⁢ ⁢ given ⁢ ⁢ by ⁢ ⁢ Equation ⁢ ⁢ ( 15 ) . ( 14 ) D 2 ⁡ ( n ) = 1 K ⁢ ∑ i = 1 K ⁢ D i 2 ⁡ ( n ) , ( 15 )

where D 2 (n) is an average of D i 2 (n), and K is the number of feature points.

According to the definition of D 2 (n), a value of D 2 (n) is the distance between feature points of two successive frames. The value of D 2 (n) is mainly related to motion information.

If an object moves faster, factors, such as motion blur and the exceeding of a search range of the tracking algorithm, are generated. These factors increase the possibility of tracking errors or tracking bias. Therefore, the value of D 2 (n) increases in such cases.

If the value of D 2 (n) is greater than a predetermined threshold value T 1 , that is, D 2 >T 1 , a serious tracking error exists. Therefore, a system is automatically restarted in order to prevent a tracking error caused by dramatic motions.

The difference between D 1 (n) and D 2 (n) is that D 1 (n) indicates an error caused by the accumulated tracking result and D 2 (n) indicates an instantaneous tracking error.

In addition, a variance of D 1 is given by Equation (16) below.

V 1 ⁡ ( n ) = 1 K ⁢ ∑ i = 1 K ⁢ D i 1 ⁡ ( n ) · D i 1 ⁡ ( n ) - D _ 1 ⁡ ( n ) · D _ 1 ⁡ ( n ) , ( 16 )

where V 1 (n) indicates a variance of D i 1 (n), K indicates the number of feature points, and indicates an average of D i 1 (n).

Since V 1 (n) is the variance of D 1 (n), ever-present external factors, such as image distortion and quantization and average tracking effects, can be reduced.

If an input object is a flat object, V 1 (n) will have a small value. On the other hand, if the input object is a 3D object, V 1 (n) will have a greater value than that of the flat object.

The liveness value calculation unit 250 calculates a liveness value using values of D 1 , D 2 and V 1 calculated by the distance calculation unit 240 and using Equation (17).

In other words, the liveness value L(n) is obtained by dividing V 1 (n) by D 1 (n) and D 2 (n).

The liveness value may also be calculated using Equation (18) in order to increase the robustness of the liveness value. That is, L(n) given by Equation (17) is divided by frames F.

If the liveness value L given by Equation (18) is greater than a predetermined threshold value T 2 , that is, L>T 2 , the liveness detection unit 260 determines that the input video image is a real object. If the liveness value L is less than the predetermined threshold value T 2 , that is, L<T 2 , the liveness detection unit 260 determines that the input video image is a photographed object. The threshold value T 2 can be determined by a user or system characteristics.

In summary, the biggest difference between images of a photographed face and a real face is that the photographed face is a flat object while the real face is a 3D object with relatively large depth changes. This difference is used to define three statistical distances D 1 , D 2 and V 1 .

D 1 indicates an accumulated tracking error. The value of D 1 increases when feature points deviate from their initial forms. D 2 is mainly related to motion information. As feature points move faster, the value of D 2 is increased.

V 1 includes 3D depth changes. If tracked feature points are on the same plane or hardly show depth changes, the value of V 1 is very small. If not, the value of V 1 is large.

As described above, the liveness value L defined in the present embodiment is obtained by dividing V 1 by D 1 and D 2 . The liveness value L reflects 3D information of a tracked object which is compensated for an accumulated tracking error and a motion tracking error.

FIGS. 4 through 6 compare statistical distances D 1 , D 2 and V 1 of a real object with those of a photographed object according to another embodiment of the present invention.

FIGS. 4 through 6 illustrate histograms of the statistical distances D 1 , D 2 and V 1 of a real face and a photographed face.

The X axis indicates values of D 1 , D 2 and V 1 , which are quantized into 256 levels, within a predetermined range, i.e., between a minimum value and a maximum value. The Y axis indicates the number of frames.

The values of D 1 and D 2 when a video image is a real face 400 or 500 and when it is a photographed face 410 or 510 are similar under general tracking conditions.

However, while the value of V 1 is large when the video image is a real face 600 , it is smaller when the video image is a photographed face 610 .

FIG. 7 is a flowchart illustrating a liveness detection method according to another embodiment of the present invention.

Referring to FIG. 7 , the liveness detection apparatus obtains a video image and determines whether a current system is in a tracking mode in operations S 700 and S 702 . The tracking mode is performed after an object and its feature point are detected in the video image.

If the current system is not in the tracking mode, the liveness detection apparatus detects an object and its feature point from a next frame in operation S 704 . After detecting the object and its feature point are performed, the liveness detection apparatus tracks feature points of subsequent frames in operation S 706 .

In operation S 708 , the detected feature point and a feature point of an (n−1) th frame are affine-transformed for an n th frame. In operation S 710 , the distance D 1 between the initially detected and affine-transformed feature point and a feature point of the n th frame is calculated, and the distance D 2 between the affine-transformed feature point of the (n−1) th frame and the feature point of the n th frame, i.e., a current frame, is calculated. In addition, the variance V 1 of the distance D 1 between the initially detected and affine-transformed feature point and the feature point of the n th frame is calculated.

In operation S 712 , a liveness value L is calculated based on the statistical distance values D 1 , D 2 and V 1 , which were calculated in operation S 710 , and using Equation (18) above.

In operation S 714 , it is determined whether the video image is a real object using the liveness value L calculated in operation S 712 .

FIG. 8 is a flowchart illustrating a liveness detection method according to another embodiment of the present invention.

Referring to FIG. 8 , an image frame is acquired from an input video image, and a face is detected in the acquired image frame in operations S 800 and 802 . In operation S 804 , corner points are detected in the detected face region. In operation S 806 , feature points are selected based on the detected corner points. In operation S 808 , feature points of subsequent frames are tracked using a tracking algorithm, e.g., the LKT algorithm. In operation S 810 , statistical distances D 1 , D 2 and V 1 for the detected feature points and the tracked feature points are calculated. In operation S 812 , it is determined whether D 2 is greater than a predetermined threshold value T 1 . This operation is performed to restart a system when an excessive tracking error occurs. In other words, if D 2 >T 1 , a video image is acquired again back in operation S 800 , and subsequent operations are also performed again.

If D 2 <T 1 , a liveness value L is calculated using Equation (18) in operation S 814 . In operation S 816 , it is determined whether the liveness value L is greater than a predetermined threshold value T 2 . If L>T 2 , it is determined that the input video image is a real face. If L<T 2 , it is determined that the input video image is a photographed face.

FIG. 9 is a schematic block diagram of an access control system using a liveness detection apparatus 870 , according to another embodiment of the present invention.

Referring to FIG. 9 , the access control system includes a camera 800 , an image acquisition unit 810 , an image control unit 820 , a face detection unit 830 , a face recognition unit 840 , an access control unit 850 , a database (DB) 860 , the liveness detection apparatus 870 , an input unit 880 , and a door control unit 890 .

The access control system detects a video image, e.g., a person's face, photographed by the camera 800 , recognizes the detected face, and opens the door. In this case, the access control system can determine whether the photographed video image is a real face or a photographed face of the person using the liveness detection apparatus 870 according to the present invention.

The present invention can also be implemented as computer-readable code on a computer-readable recording medium. The computer-readable recording medium is any data storage device that can store data which can be thereafter read by a computer system. Examples of the computer-readable recording medium include read-only memory (ROM), random-access memory (RAM), CD-ROMs, magnetic tapes, floppy disks, optical data storage devices, and carrier waves (such as data transmission through the Internet).

The computer-readable recording medium can also be distributed over network-coupled computer systems so that the computer-readable code is stored and executed in a distributed fashion. Also, functional programs, code, and code segments for accomplishing the present invention can be easily construed by programmers skilled in the art to which the present invention pertains.

According to the present invention, object feature points are tracked from a video image, and detected feature points are affine-transformed. Then, it is determined whether the video image is a real object using characteristics of the affine-transformed feature points. Therefore, liveness detection resistant to spoofing can be performed without requiring user intervention and additional hardware.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it will be understood by those of ordinary skill in the art that various changes in form and details may be made therein without departing from the spirit and scope of the present invention as defined by the following claims. The exemplary embodiments should be considered in descriptive sense only and not for purposes of limitation. Therefore, the scope of the invention is defined not by the detailed description of the invention but by the appended claims, and all differences within the scope will be construed as being included in the present invention.