CN105023010B - A kind of human face in-vivo detection method and system - Google Patents

Abstract

The invention discloses a human face living body detection method and system, wherein the human face living body detection method includes the steps of: simultaneously acquiring two images of the recognition object through two cameras; using a face classifier to locate the two images The face area of the face is obtained, and the face images corresponding to the two images are respectively obtained; the feature points of the two face images are located; the feature points of the face image after the feature point positioning are used for fast stereo matching of the feature points; the two face images are calculated In the disparity of the matching feature points, the depth information value of the matching feature point is obtained; according to the depth information values of the multiple matching feature points, it is judged whether the face of the recognition object is a living body. Adopting the present invention does not require too much cooperation from users, has high security, strong concealment, strong posture adaptability, wide adaptability, fast matching speed, and good user experience.

Description

Method and system for human face liveness detection

technical field

The invention relates to the technical field of pattern recognition, in particular to a method and system for detecting human face liveness.

Background technique

With the development of biometric technology, face recognition technology has become mature. At present, under good lighting conditions and postures, the face recognition system can already perform relatively accurate face detection and recognition. However, for some For the face recognition system of access control, login and other systems, users can deceive the system by illegal means such as photos. Therefore, for this type of system, the higher the recognition rate, the greater the potential safety hazard. Therefore, the realization of living body detection has become a means of guaranteeing the security of this type of system.

In the prior art, some face detection methods used in face recognition systems require the user to cooperate with specific actions. The algorithm is simple, but it is easy to leak the judgment basis, making it easy for illegal users to imitate, and it takes a long time to cooperate with the action. Limited, some methods such as blink detection and optical flow detection generally require the user to accurately face the camera, the posture adaptability is not high, and the user experience is not good.

Therefore, how to overcome the above shortcomings is a technical problem to be solved urgently by those skilled in the art.

Contents of the invention

In view of this, the present invention aims to provide a face detection method and system, which not only have strong adaptability to gestures, high security, but also do not require user cooperation, are fast, and provide good user experience.

Specifically, described human face detection method comprises steps:

S1. Simultaneously acquire two images of the recognition object through two cameras respectively;

S2. Use a face classifier to locate the face areas in the two images, and obtain face images respectively corresponding to the two images;

S3, performing feature point positioning on the two face images;

S4. Perform fast stereo matching of feature points on the face image after feature point positioning;

S5. Calculate the parallax of the two face images at the matching feature points, and obtain the depth information value of the matching feature points;

S6. According to the depth information values of the plurality of matching feature points, it is judged whether the human face of the recognition object is a living body.

Preferably, in the embodiment of the present invention, the two cameras are arranged left and right or up and down to form a binocular stereo vision system.

Preferably, in the embodiment of the present invention, the face classifier is based on at least one of the following methods: skin color and geometric characteristics, integral projection, template matching, and line connection.

Preferably, in the embodiment of the present invention, the feature points include all or part of the key points located in at least one of the following regions: human face outline, eyebrows, eyes, nose and mouth.

Preferably, in the embodiment of the present invention, in step S3, at least one of the following algorithms is used for the positioning of feature points: deep learning related algorithms, active shape model related algorithms, active appearance model related algorithms, and cascaded shape regression related algorithms .

Preferably, in the embodiment of the present invention, step S4 specifically includes: taking a feature point in one face image as a reference feature point, and acquiring the same reference feature point around the corresponding feature point of another face image by means of accumulated absolute error (SAD) Corresponding matching feature points.

Preferably, in the embodiment of the present invention, the ordinates of the reference feature point and the matching feature point are the same; or, when the ordinates of the reference feature point and the corresponding matching feature point are different, the matching feature point The ordinate of the point is forced to be equal to the ordinate of the corresponding reference feature point, and the matching point is obtained, and two pixel points are taken around the matching point, and the matching point and each pixel point are used as the candidate points of the matching feature point, and are calculated separately The error accumulation value centered on the candidate point, and the candidate point with the smallest error accumulation value is selected as the matching feature point.

Preferably, in the embodiment of the present invention, step S6 adopts at least one of the following methods: support vector machine SVM algorithm, feature extraction algorithm based on principal component analysis PCA, linear discriminant analysis LDA algorithm; or, step S6 includes:

S61. Determine the matching feature point with the smallest depth information value as the smallest matching feature point, and acquire the depth information value of the smallest matching feature point;

S62. Subtract the depth information value of the minimum matching feature point from the depth information value of each matching feature point to obtain the relative depth information value of each matching feature point;

S63. Calculate the sum or square sum of the relative depth information values of each matching feature point and use it as a comparison value. When the comparison value is smaller than a preset threshold, it is determined that the recognition object is not a living body, otherwise it is determined as a living body.

Preferably, in the embodiment of the present invention, after step S63, it also includes:

S64, for the identification object determined to be a living body through step S63, according to the feature point positioning result, estimate the size of the face, calculate the face aspect ratio of the left and right two face images, when the face width and height of the left and right two face images When the ratio exceeds the preset interval, it is determined that the identification object is not a living body.

Preferably, in the embodiment of the present invention, the method further includes a step before step S1: S0, calibrating the binocular stereo vision system composed of the two cameras.

Preferably, in the embodiment of the present invention, the method further includes a step before step S2: performing image preprocessing operations including stereo correction on the two images.

Preferably, in the embodiment of the present invention, the stereoscopic correction of the image includes: reprojecting the image planes of the two cameras according to the calibration result in step S1, so that the two images fall on the same plane precisely, And the rows of the two images are perfectly aligned on the front-parallel structure.

On the other side of the embodiments of the present invention, a human face liveness detection system is also provided, the system comprising:

Two cameras, used to simultaneously acquire two images of the identified object respectively;

The face classifier is used to locate the face area in the two images acquired by the two cameras, and obtain the face images corresponding to the two images respectively;

A feature point positioning unit is used to perform feature point positioning on two face images;

Stereo matching unit, for performing fast stereo matching of feature points on the face image after feature point positioning;

A calculation unit is used to calculate the parallax of the two face images at the matching feature points, and obtain the depth information value of the matching feature points;

The processing unit is configured to judge whether the face of the recognition object is a living body according to the depth information values of the plurality of matching feature points.

In at least one solution of the present invention, dual cameras are used to simulate human eyes, and two or more two-dimensional images of the user (recognition object) are captured by the two cameras from different angles, and finally converted into world coordinates through a series of techniques The three-dimensional physical information in the system, in which the internal and external parameters of the camera itself are determined through the calibration of the camera, and then the captured two-dimensional plane images are processed, so that the captured images can be located on the same plane, and the pixel points of the left and right images Perform stereo matching, mark the points in the three-dimensional object in two two-dimensional images, and then place the model in the affine space according to the parallax principle of the bionic human eye for calculation and restoration to obtain the depth information value of the object.

When the binocular stereo vision system is used to photograph the user, the obtained facial organs such as the nose and the depth of the eyes are obviously lower than the contour of the face, and the depth difference between the facial organs and the contour of a certain side is more obvious when the head posture changes. Videos cannot show discriminative depth differences. Even if a large-sized image can show obvious depth differences in the case of curvature, it will inevitably sacrifice the normal face ratio. By calculating the depth information value and prior knowledge of different feature points in the face image, it can be judged whether the user is alive or not.

It can be seen that, after adopting the solution of the present invention, in the detection of human face liveness, there is no need for too much cooperation from the user, and the security is high, the concealment is strong, the posture adaptability is strong, the adaptability is wide, and the feature points are used in image matching. The positioning technology can accurately find out the position of the matching feature point, and only need to search in a small area to achieve the image matching of the feature point, which greatly simplifies the image matching process, improves the matching speed, and provides a good user experience.

Description of drawings

The drawings constituting a part of the present invention are used to provide a further understanding of the present invention, and the schematic embodiments and descriptions of the present invention are used to explain the present invention, and do not constitute an improper limitation of the present invention. In the attached picture:

FIG. 1 is a schematic diagram of the steps of the face detection method described in the embodiment of the present invention;

Fig. 2a is a schematic diagram of the left and right images before stereo correction in the face liveness detection method according to the embodiment of the present invention;

Fig. 2b is a schematic diagram of the left and right images after stereo correction in the face liveness detection method according to the embodiment of the present invention;

Fig. 3 is a feature point positioning result and a matching feature point position diagram in the right image in the human face living body detection method according to an embodiment of the present invention;

Fig. 4a is a schematic diagram of the feature point depth information value of a real face in the face liveness detection method according to an embodiment of the present invention;

Fig. 4b is a schematic diagram of depth information values of feature points of face photos in the face liveness detection method according to the embodiment of the present invention;

FIG. 5 is a schematic structural diagram of a human face liveness detection system provided by an embodiment of the present invention.

Detailed ways

It should be pointed out that the description and sequence of specific structures in this section are only descriptions of specific embodiments, and should not be considered as limiting the protection scope of the present invention. In addition, the embodiments in this section and the features in the embodiments can be combined with each other under the condition of no conflict.

Please refer to FIG. 1 to FIG. 5 at the same time. The human face liveness detection method and system according to the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

As shown in Figure 1, the human face activity detection method of the present invention may include:

Step 1: Use two cameras to simultaneously capture two images of the identified object (the user is used as an example below), namely the left and right images;

Step 2: Use a face classifier to locate the face areas in the two images, and obtain face images corresponding to the two images respectively;

Step 3: Perform feature point positioning on two face images;

Step 4: Perform fast stereo matching of feature points on the face image after feature point positioning;

Step 5: Calculate the parallax of the two face images at the matching feature points, and obtain the depth information value of the matching feature points;

Step 6: According to the depth information values of multiple matching feature points, determine whether the user's face is a living body.

Wherein, during specific implementation, the two cameras can adopt the same type of 5 million-pixel network camera, and the two cameras are fixedly arranged on the left and right, and make a binocular stereoscopic vision system composed of the two cameras; in practical applications, between the two cameras The horizontal distance between the two cameras is set so that the two cameras are on the same plane and can acquire user images at the same time. The distance between the camera and the face is about 0.5 meters. In the general indoor lighting environment, the size of the left and right images is 640×480. Functions such as face detection, feature point positioning, living body detection, and recognition systems can be realized through a PC. .

As a preferred method, the method of the embodiment of the present invention may also include a step in step 1: calibrating the binocular stereo vision system composed of two cameras, and its specific implementation process may be as follows:

Step 11: Calibrate the two cameras respectively, and the calibration content may specifically include: internal parameter matrices and distortion vectors of the left and right cameras.

Step 12: Calibrate the binocular stereo vision system composed of two cameras. The calibration content may specifically include: the rotation matrix and the translation vector of the binocular stereo vision system.

Step 13: Obtain stereo correction parameters and a reprojection matrix.

In this way, step 11 can use the checkerboard calibration method to calibrate the camera, and the internal parameter matrix can include the focal length in the horizontal direction of the camera, the focal length in the vertical direction, and the position of the principal point; the distortion vector can be composed of radial distortion coefficients and tangential distortion coefficients.

In this way, for the calibration result obtained in step 11, the lens distortion in the radial and tangential directions can be eliminated by mathematical methods, so as to output an undistorted image, let (x _p , y _p ) be the position of the point without distortion, ( x _d , y _d ) is the distortion position, there are:

This results in an image without lens distortion.

In this way, the rotation matrix and translation vector described in step 12 can be used to describe the positional relationship of the right camera relative to the left camera, and the mathematical expression is X _R =R*X _L +T, where X _L and X _R are three-dimensional The three-dimensional position vector of any point P in the space in the reference coordinate system of the left and right cameras of the binocular stereo vision system, and R and T are the rotation matrix and translation vector of the binocular stereo vision system.

In this way, step 13 can use the Bouguet algorithm to obtain the stereo correction parameters of the binocular stereo vision system, and use the reverse mapping method to obtain the correction mapping table of the left and right views, and obtain the reprojection for reprojecting the two-dimensional image points into three dimensions. projection matrix.

In addition, before using the face classifier described in step 2 to locate the face areas in the two images, the method in the embodiment of the present invention may further include the step of: performing image preprocessing including stereo correction on the two images operate. Among them, the stereo correction operation can be based on the aforementioned calibration results, by looking up the left and right correction mapping tables, and reprojecting the image planes of the two cameras, so that the left and right images fall on the same plane precisely, and the lines of the two images are completely aligned For the front-parallel structure, that is, the pixel row of one camera is on the same row as the other at the same point, the left and right images before and after stereo correction can be shown in Figure 2a and Figure 2b.

In this way, the face classifier is used to locate the face areas in the two images as described in step 2, and the image preprocessing operation adopted can be performed as grayscale transformation and filtering processing on the stereo-corrected image to obtain the previous parallel-aligned and high-quality gray-scale images; the face location method adopted can be a method based on template matching, that is, the respective Haar-Like wavelet eigenvalues are quickly calculated by integrating the left and right gray-scale images, and the application Go to the Adaboost-Cascade classifier trained offline to detect the face area of the left and right images. Of course, positioning can also be realized based on skin color and geometric characteristics, based on integral projection, based on line connectivity and other methods.

In this mode, the feature point location is carried out to the face image obtained by positioning described in step 3, and the feature point refers to the position coordinates of the outline of the face, eyebrows, eyes, nose and mouth; in this embodiment, cascading is used The feature point location is performed by the shape regression algorithm. Of course, in other embodiments, it can also be implemented by using a deep learning-based algorithm, an active shape model-related algorithm, an active appearance model-related algorithm, and the like. Using the cascade shape regression algorithm for feature point location, such as the two-layer cascade shape regression algorithm using index features, can directly learn a vector regression function, combine the image itself and minimize the training shape alignment error to estimate the shape of the face; combined Two-layer cascade regression, shape index feature and correlation-based feature selection method can quickly and accurately train a non-parametric regression model with extremely fast speed. It is an efficient and high-precision precise positioning algorithm for facial feature points. The specific steps are:

Step 31: Load the offline-trained two-layer cascaded shape regressor based on index features;

Step 32: Randomly select L training sample shapes from the loaded regressor as the initial shape of the feature points of the face region to be located;

Step 33: Calculate the estimated user's face shape obtained when a certain selected training sample is used as the initial shape;

Step 34: Take the average value of the L estimated user's face shapes as the user's face shape.

Among them, the shape refers to the vector composed of the position information of the feature points of the face, and the mathematical expression is S={x ₁ , y ₁ , x ₂ , y ₂ ,..., x _n , y _n }, x _i , y _i are the ith The pixel coordinates corresponding to the feature points.

In this way, step 33 specifically includes the following processes:

Step 331: Calculating F grayscale difference features;

Step 332: Obtain the update value of the current shape, and update the current shape;

Step 333: Complete the shape update for the prescribed number of regressions to obtain the final estimated shape of the current user's face.

Wherein, the acquisition method of the gray scale difference feature described in step 331 may be: according to the F pair of feature point index numbers and relative position information in the regressor, obtain the gray scale value of the corresponding index feature on the user's face image, and calculate each pair of index Gray difference of feature points.

The method for obtaining the update value of the current shape described in step 332 may be: compare the F gray difference values obtained in step 331 with the corresponding thresholds to obtain the corresponding F-bit secondary value, and then find out the The updated shape corresponding to this binary value.

The specific algorithm for updating the current shape described in step 332 can be:

S _i =S _i-1 +δS

Among them, S _i-1 is the previous estimated shape of the current estimated shape, and 6S refers to the updated shape.

In addition, the rapid stereo matching of the feature points on the face image after the feature point positioning described in step 4 means that the small window of "Sum of Absolute Difference (SAD, Sum of Absolute Difference)" can be used to find the difference between the left and right stereo corrected images. matching feature points. Using each feature point in the left image as a reference point, use the SAD sliding window to perform matching search around the corresponding feature point in the right image. The point with the smallest accumulated absolute error is the matching feature point of the left image feature point on the right image.

In this manner, the accumulated absolute error value of the position of the window during the search and matching process of the SAD sliding window can be:

Since step 3 can firstly carry out accurate facial feature point positioning on the left and right images, and the image size is 640×480, so a predetermined number of SAD sliding windows with a predetermined size (such as 5) are used, and a predetermined number of points are centered on the right image feature point. Search for matches within a small window of (eg 5) pixels.

In this way, the matching feature points after the fast stereo matching of the feature points described in step 4 can be required to have the same vertical coordinates. When the vertical coordinates of the matching feature points are different, the vertical coordinates of the matching feature points in the right image can be forced to be equal to the The image corresponds to the ordinate of the referenced feature point, and the matching point M' is obtained, and two pixel points are taken on the left and right sides of M', and these 5 points are used as candidate points of M', and the error centered on the candidate point is calculated respectively Cumulative value, select the point with the smallest error cumulative value as the matching point. Attached figure 3 is the positional relationship between the feature point positioning result and the matching feature point in the right image. It can be seen from the figure that the matching feature point is near the positioned feature point. Saves matching time.

Additionally, step 5 may include:

Step 51: Calculate the parallax of the left and right face images at the matching feature points;

Step 52: Obtain the depth information value of the feature point according to the parallax of the feature point.

Wherein, the parallax in step 51 may be the abscissa distance of the left and right matching feature points.

The depth information value of the matching feature point in step 52 can be obtained by using the principle of similar triangles, and the calculation method of the depth information value of the matching feature point is Among them, T refers to the distance between the projection centers of the two cameras, and f is the focal length of the two cameras of the same model. In the present invention, two cameras of the same model are used, T and f are calibrated in advance, and d is the disparity of left and right matching feature points. Knowing the disparity d and the two-dimensional image point (x, y), the reprojection matrix Q in step 1 can be used to calculate the three-dimensional coordinates corresponding to the two-dimensional image point in the image:

is the three-dimensional coordinates of the point, is the depth of the 2D image point.

In addition, as described in step 6, according to the depth information value of matching feature points in different parts, it is judged whether the face is a living body, and the specific steps can be as follows:

Step 61: Determine the matching feature point with the smallest depth information value as the minimum matching feature point, and obtain the depth information value of the minimum matching feature point;

Step 62: respectively subtracting the depth information value of the minimum matching feature point from the depth information value of each matching feature point to obtain the relative depth information value of each matching feature point;

Step 63: Calculate the sum of squares or the sum of depths of the relative depth information values of each matching feature point, and compare it with a preset threshold, and if the sum of squares is less than the threshold, it is determined to be non-living; otherwise, it is determined to be living;

When implementing step 6, the matching feature point depth information values obtained by using the user's real face (living body) and the user's face photo as identification objects can be shown in Figure 4a and Figure 4b, wherein Figure 4a is A schematic diagram of the feature point depth information value of a real face in the face liveness detection method of the embodiment of the present invention. FIG. 4b is a schematic diagram of the face photo feature point depth information value in the face liveness detection method of the embodiment of the present invention. According to the comparison of Figure 4a and Figure 4b, it can be seen that the depth information value between different matching feature points obtained by using the user's real face as the identification object changes significantly, while the depth information value obtained by using the user's face photo as the identification object The change of the depth information value between different matching feature points is relatively inconspicuous.

Step 64: Estimate the size of the face according to the feature point positioning results, and calculate the face aspect ratio of the left and right face images. When the face aspect ratios of the left and right face images exceed the preset interval, determine the The above-mentioned identification object is a non-living body.

The purpose of step 64 is to further identify the identification object that has been determined to be living in step 63, so as to improve the identification rate and avoid fraudulent authentication.

In practical applications, it is possible to pass the determination of step 63 by bending a large photo (close to the size of a real human face), that is, step 63 may be fraudulently authenticated. In order to avoid this fraudulent authentication, in the embodiment of the present invention, a judging step of the face aspect ratio is also included, specifically:

For the user who has been judged to be living in step 63, in this step, the face aspect ratios of the left and right face images are calculated respectively, and whether the recognition object is judged by calculating whether the face aspect ratio exceeds the preset interval For bending large photos.

In practical applications, after the first and second judgments that the height and width of the face exceed the preset interval, return to step 1 to re-acquire the image judgment, and when the three times (or other predetermined times) are greater than the threshold, it is judged as non-living , so as to avoid misjudgment through multiple judgments. This constrained approach prevents crooked authentication by bending large photos. Then, if it is finally determined to be a living body, it will enter the face recognition link.

In addition, in the whole process, the binocular stereo vision system can be used to take 200 photos of the real user and 200 face photos respectively. Operations such as bending, and experiments using 6-inch and 12-inch face photos can completely and correctly distinguish whether the user is alive or not.

It should be noted that the solutions of the above-mentioned embodiments are not only applicable to binocular stereo vision systems with cameras fixedly arranged left and right, but also applicable to binocular stereo vision systems with fixed arrangements up and down (two cameras are arranged up and down, and the distance between the two cameras The vertical distance is set so that the two cameras are on the same plane and can acquire user images at the same time). For the solution of adopting the vertical arrangement, the aforementioned left image corresponds to the upper image in the vertical arrangement structure, the right image corresponds to the lower image, the abscissa corresponds to the ordinate, and the ordinate corresponds to the abscissa.

It should be noted that the classification method for judging whether a face is alive or not according to the depth information value of matching feature points in different parts described in step 6 is not limited to the above-mentioned scheme, and support vector machine SVM and PCA based on principal component analysis can also be used. Feature extraction, linear discriminant analysis LDA and other methods, as long as the corresponding functions can be realized.

It should be noted that, the image preprocessing operation may select different preprocessing methods according to different modes of the face classifier.

After the face detection method of the embodiment of the present invention is applied, two cameras can be used to simulate the eyes of a person, and the two cameras can take two or more two-dimensional images of the user from different angles, and finally convert them into world coordinates through a series of technologies The three-dimensional physical information in the system, in which the internal and external parameters of the camera itself are determined through the calibration of the camera, and then the captured two-dimensional plane images are processed, so that the captured images can be located on the same plane, and the pixel points of the left and right images Perform stereo matching, mark the points in the three-dimensional object in two two-dimensional images, and then place the model in the affine space according to the parallax principle of the bionic human eye for calculation and restoration to obtain the depth information value of the object.

When the binocular stereo vision system is used to photograph the user, the obtained facial organs such as the nose and the depth of the eyes are obviously lower than the contour of the face, and the depth difference between the facial organs and the contour of a certain side is more obvious when the head posture changes. Videos cannot show discriminative depth differences. Even if a large-sized image can show obvious depth differences in the case of curvature, it will inevitably sacrifice the normal face ratio. By calculating the depth information value and prior knowledge of different matching feature points in the face image, it can be judged whether the user is alive or not. It can be seen from the above that, compared with the prior art, the solution of the embodiment of the present invention not only has strong adaptability to gestures and high security, but also requires no user cooperation, is fast, and provides good user experience.

In addition, as shown in FIG. 5 , an embodiment of the present invention also provides a human face detection system, which may include two cameras, a face classifier, a feature point positioning unit, a stereo matching unit, a computing unit and a processing unit . Among them: two cameras are used to obtain the user's image respectively; the face classifier is used to locate the two images acquired by the two cameras into the face area; the feature point positioning unit is used to locate the face image obtained by the face classifier Perform feature point positioning; the stereo matching unit is used to perform fast stereo matching of feature points on the face image after feature point positioning; the calculation unit is used to calculate the depth information value of the matching feature point in the face image; the processing unit is used for according to different parts Match the depth information value of the feature points to determine whether the user's face is alive.

During specific implementation, the functions of the above-mentioned units and the face classifier are not limited to be realized by a PC, as long as the corresponding functions of these parts can be realized. In addition, for other extensions and descriptions of the human face liveness detection system, reference may be made to the relevant descriptions of the method embodiments, which are not repeated here; and the system can also achieve corresponding technical effects.

Those of ordinary skill in the art can understand that some of the steps/units/modules of the above embodiments can be implemented by hardware related to program instructions, and the aforementioned programs can be stored in computer-readable storage media. When the program is executed, the execution includes The corresponding steps in each unit of the above-mentioned embodiment; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk and other various media that can store program codes.

The specific embodiments described above have further described the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above descriptions are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (13)

1. a human face living body detection method, is characterized in that, described method comprises steps: S1. Simultaneously acquire two images of the recognition object through two cameras respectively; S2. Use a face classifier to locate the face areas in the two images, and obtain face images respectively corresponding to the two images; S3, performing feature point positioning on the two face images; S4. Perform fast stereo matching of feature points on the face image after feature point positioning; S5. Calculate the parallax of the two face images at the matching feature points, and obtain the depth information value of the matching feature points; S6. According to the depth information values of a plurality of matching feature points, it is judged whether the human face of the recognition object is a living body; Step S6 includes: S61. Determine the matching feature point with the smallest depth information value as the smallest matching feature point, and acquire the depth information value of the smallest matching feature point; S62. Subtract the depth information value of the minimum matching feature point from the depth information value of each matching feature point to obtain the relative depth information value of each matching feature point; S63. Calculate the sum or square sum of the relative depth information values of each matching feature point and use it as a comparison value. When the comparison value is smaller than a preset threshold, it is determined that the recognition object is not a living body, otherwise it is determined as a living body. 2. the human face living body detection method as claimed in claim 1, is characterized in that, described two cameras form binocular stereo vision system by arranging left and right or arranging up and down. 3. The human face liveness detection method according to claim 1, wherein the human face classifier is based on at least one of the following methods: skin color and geometric characteristics, integral projection, template matching, and line connection. 4. The human face liveness detection method according to claim 1, wherein the feature points include all or part of the key points located in at least one of the following areas: human face outline, eyebrows, eyes, nose and mouth. 5. The face detection method according to claim 1, characterized in that, in step S3, at least one of the following algorithms is used for the positioning of feature points: deep learning related algorithms, active shape model related algorithms, active appearance models Correlation Algorithms and Cascaded Shape Regression Correlation Algorithms. 6. The human face living body detection method as claimed in claim 1, wherein step S4 is specifically: taking a feature point in a face image as a reference feature point, corresponding to another face image by means of absolute error accumulation SAD The matching feature points corresponding to the reference feature point are acquired around the feature point. 7. human face living body detection method as claimed in claim 6, is characterized in that, the ordinate of described reference feature point and described matching feature point is identical; Or, when the ordinate of reference feature point and corresponding matching feature point When not the same, the ordinate of the matching feature point is forced to be equal to the ordinate of the corresponding reference feature point to obtain the matching point, and two pixel points are respectively taken around the matching point, and the matching point and each pixel point are used as matching points. Candidate points of feature points, and respectively calculate the cumulative error value centered on the candidate point, and select the candidate point with the smallest cumulative error value as the matching feature point. 8. human face detection method as claimed in claim 1, is characterized in that, step S6 adopts at least one of the following modes: support vector machine SVM algorithm, feature extraction algorithm based on principal component analysis PCA, linear discriminant analysis LDA algorithm. 9. human face detection method as claimed in claim 8, is characterized in that, also comprises after step S63: S64, for the identification object determined to be a living body through step S63, according to the feature point positioning result, estimate the size of the face, calculate the face aspect ratio of the left and right two face images, when the face width and height of the left and right two face images When the ratio exceeds the preset interval, it is determined that the identification object is not a living body. 10. The human face living body detection method according to any one of claims 2 to 9, characterized in that, the method also includes the step before step S1: S0, the binocular stereo vision system formed by the two cameras Calibrate. 11. The human face liveness detection method according to claim 10, characterized in that, before step S2, the method further comprises the step of: performing image preprocessing operations including stereo correction on the two images. 12. The human face liveness detection method according to claim 11, wherein the stereoscopic correction of the image comprises: reprojecting the image planes of the two cameras according to the calibration result in step S1, so that the two images fall exactly on the same plane, and the lines of the two images are perfectly aligned on the front-parallel structure. 13. A human face living body detection system, comprising: Two cameras, used to simultaneously acquire two images of the identified object respectively; The face classifier is used to locate the face area in the two images acquired by the two cameras, and obtain the face images corresponding to the two images respectively; A feature point positioning unit is used to perform feature point positioning on two face images; Stereo matching unit, for performing fast stereo matching of feature points on the face image after feature point positioning; A calculation unit is used to calculate the parallax of the two face images at the matching feature points, and obtain the depth information value of the matching feature points; A processing unit, configured to determine whether the face of the recognition object is a living body according to the depth information values of a plurality of matching feature points; The processing unit is also specifically used for: Determining the matching feature point with the smallest depth information value as the minimum matching feature point, and obtaining the depth information value of the minimum matching feature point; Respectively subtracting the depth information value of the minimum matching feature point from the depth information value of each matching feature point to obtain the relative depth information value of each matching feature point; The sum or the square sum of the relative depth information values of each matching feature point is calculated and used as a comparison value. When the comparison value is less than a preset threshold, it is determined that the recognition object is not a living body, otherwise it is determined as a living body.