CN104036229A - Regression-based active appearance model initialization method - Google Patents

Abstract

The invention discloses a regression-based active appearance model initialization method, which belongs to the field of computer vision. The implementation process of this method is as follows: when using the active appearance model for automatic tracking of facial feature points, assuming that the target position information of the first frame in the video tracking is known, in the subsequent tracking process, use the double-threshold feature correspondence algorithm to obtain adjacent The discrete feature correspondence between frame images uses the spatial mapping relationship between discrete feature points and structured calibration points established by the Kernel Ridge regression algorithm to obtain the initial calibration of facial features, which can greatly reduce the number of subsequent iterations and improve calibration accuracy. Compared with the initialization method of the traditional active appearance model, the method of the present invention can assist the active appearance model to obtain more accurate calibration results of facial feature points.

Description

Regression-Based Active Appearance Model Initialization Method

technical field

The invention belongs to the technical field of image analysis, in particular to a regression-based active appearance model initialization method.

Background technique

In the field of computer vision research, the use of Active Appearance Model (Active Appearance Model, AAM) method to locate target shape feature points is a hot spot of attention and research in recent years. It was first proposed by Edwards et al. It has been widely used in the registration and recognition of other non-rigid bodies. The AAM algorithm is an improvement to the Active Shape Model (ASM). Compared with ASM, it considers the constraints of global information and adopts the statistical constraints of shape and texture fusion, that is, the statistical appearance constraints. Moreover, the search principle of AAM draws on the main idea of analysis-by-synthesis (ABS), and makes the model gradually approach the actual input model by continuously adjusting the parameters in the model.

Although AAM is effective, no matter what kind of improved algorithm it is, the fitting efficiency and accuracy of the algorithm are closely related to the given initial position of the model. If the initial position is poor, the fitting algorithm based on gradient descent in AAM will most likely fall into a local minimum and obtain very poor target shape positioning results. Therefore, the determination of the initial feature points is the key factor affecting the robustness and speed of the algorithm, and the automatic and accurate calibration of the face feature points can greatly improve the efficiency and accuracy of the algorithm. The traditional AAM initialization methods include: (1) using the average face as the initial shape, and the initialization error is large; (2) using face features (such as eyes, mouth) positioning information to complete the initialization, but there are high requirements for positioning accuracy; (3) In video tracking, the localization result of the previous frame is used as the initial information of the current frame localization, but it can only adapt to the case of small changes between frames.

Contents of the invention

The object of the present invention is to propose a regression-based active appearance model initialization method based on the shortcomings of existing active appearance model initialization methods, taking video tracking of facial feature points as the research object. Because the method considers the correlation between frames, the invention can greatly improve the tracking performance of the target shape in the tracking environment of frame loss.

The technical scheme solved by the present invention is: the algorithm assumes that the key feature points of the first frame are known, firstly, the double threshold method is used to obtain the accurate local feature points corresponding to the previous frame and the current frame image, and then the Kernel Ridge regression algorithm (Kernel Ridge Regression Algorithm (Kernel Ridge Regression) Ridge Regression (KRR) establishes the mapping relationship between the discrete local feature point correspondence and the structured calibration point, and extracts the calibration point positioning information of the current frame from the local feature correspondence. The specific implementation steps of the technical solution of the present invention are as follows:

1. Select the training video, and use the kernel ridge regression algorithm to establish the mapping M _v of the spatial position between the scattered local feature points and the facial feature calibration points;

2. Use the Cascade Adaboost algorithm to detect faces, and normalize the face image to a size of 250*250;

3. Calculate the error between the reconstructed image of the previous frame and the current reconstructed image Where I ₁ and I ₂ are the face image of the previous frame and the current face image respectively, x is the pixel set under the mean shape _s0 , p is the deformation parameter from the mean shape _s0 to the current reconstructed shape s, W(x; p) is the set of pixels under the reconstruction shape s, and N is the number of pixels under the mean shape. When e>e ₀ , it means that the face image in the previous frame is quite different from the current face image, and go to step (3); otherwise, go to step (5), where e ₀ =5e-4 is the error threshold.

4. Extract the SIFT features of the previous frame face image I ₁ and the current frame face image I ₂ , use the feature matching method based on double thresholds to match, and obtain the matching pair C={(c _k , c′ _k ), k= 1, 2, ..., q _C }, where q _C is the number of matching pairs;

5. Extract a space vector V={V _k , k=1, 2, . . . , n} from the face image I ₁ of the previous frame, where n is the number of face feature points.

6. According to the mapping M _v parameters obtained in step (1) and the matching points obtained in step (4), establish the spatial position vector V of discrete feature points in the test stage and send it to the mapping M _v as input, and output the corresponding person Face calibration point, that is, to obtain the initial value of the active appearance model used for current frame tracking;

The above-mentioned regression-based active appearance model initialization method, the specific implementation process in step 1 is as follows:

(1) Data initialization, let time k=0;

(2) Between the previous frame face image I ₁ (k) and the current face image I ₂ (k), scattered matching points are obtained by establishing a matching method of an equalized probability model, so that k=k+1;

(3) Acquire KRR training data from the previous frame face image and the current face image according to the scattered matching points $T_{V_j}\left(k\right)=\{(V_k^j,V_{k^{\′}}^j),j=\mathrm{1,2},...,m\},$ Among them, m=66 is the number of calibration points;

(4) If k<n, return to (1.2), otherwise get the total training data Where n=100 is the logarithm of training samples.

(5) For each face calibration point j, first according to the training data Calculate the kernel function matrix K _j , where $K_j(V_{k_1}^j,V_{k_2}^j)=\exp (-{\left|\right|V_{k_1}^j-V_{k_2}^j\left|\right|}^2/\mathrm{\&sigma;}),k_1=\mathrm{1,2},...\mathrm{no},k_2==\mathrm{1,2},...\mathrm{no},$ where σ=0.025; then create an identity matrix I of the same size as the matrix K _j , where I(k ₁ , k ₂ )=1, k ₁ =1,2,...n,k ₂ ==1,2, ...n; then calculate the kernel coefficient vector α _j , where where λ=0.5×10 ^-7 ; finally, according to the above calculation, the regression kernel function is obtained

(6) Obtain the mapping set M _v ={f ^j (V), j=1, 2, . . . , m}.

The above-mentioned regression-based active appearance model initialization method, the specific implementation process of step 4 is as follows

(1) Double threshold initialization: set the threshold initial value η ₁ =1.5, η ₂ =8; the number of cycles iter ₁ =0, iter ₂ =0; the number of cycles iter_max ₁ =10, iter_max ₂ =20, adjacent matching points number t=5;

(2) Extract SIFT feature from previous frame face image I ₁ and current face image I ₂ ;

(3) Utilize threshold η ₁ to do SIFT feature matching, and make iter ₁ =iter ₁ +1;

(4) If the number of matches is less than t, and iter ₁ <iter_max ₁ , set η ₁ =η ₁ +0.15, and return to (4.3); if the number of matches is greater than t, or iter ₁ >iter_max ₁ , then transfer to (4.5);

(5) Utilize threshold η ₂ to do SIFT feature matching, and make iter ₂ =iter ₂ +1;

(6) If the number of matches is less than 2, and iter ₂ <iter_max ₂ , set η ₂ =η ₂ +0.02, and return (4.5); if the number of matches is greater than 2 and less than 5, and iter ₂ <iter_max ₂ , Then set η ₂ =η ₂ -0.01, and return to (4.5); otherwise, turn to (4.7);

( ₇ ) Obtain the dense matching set B={(b _j , b′ _j ), j=1, 2, ..., q _B } obtained by using η 1, where q _B is the number of dense matching, and using η ₂ The obtained exact matching set A={(a _i , a' _i ), i=1, 2, ..., q _A }, where q _A is the number of dense matches;

(8) Calculate the distance ratio of the first and last two pairs of matching points in the exact matching set

(9) For each pair of matches (b _j , b′ _j ) in the dense matching set, first calculate then calculate and get Then calculate θ=atan((a _t (y)-b _j (y))/(a _t (x)-b _j (x))), θ'=atan((a' _t (y)-b' _j (y))/(a _t (x)-b _j (x))), and calculate sig _x = sign(a _t (y)-b _j (y)), sig _y = sign(a _t (x )-b _j (x)), sig' _x = sign(a' _t (y)-b' _j (y)), sig _y = sign(a' _t (x)-b' _j (x)). if then save the current match, otherwise delete the match. Finally, the final matching pair C={(c _k , c′ _k ), k=1, 2, . . . , q _C } is obtained, where q _C is the number of final matching.

In the above-mentioned regression-based active appearance model initialization method, sub-step (3) in step 1 is carried out as follows:

(1) Suppose j is the current calibration point in I ₁ (k), o is the center of the current t adjacent matching points in I ₁ (k), and o′ is the current t matching points in I ₂ (k) corresponding to it center;

(2) Calculate the distance from the matching point i (the value of i is the number of matching points) to the center point o of the matching point and the angle between the straight line oi and the x-axis (d _i , θ _i ), and from the matching point The distance from i' to the center point o' of the matching point and the angle between the straight line o'i' and the x-axis (d' _i , θ' _i );

(3) Calculate the distance from the calibration point j to the center point o in I ₁ (k) and the angle between the straight line oj and the x-axis (d _l , θ _l );

(4) Calculate the distance from the calibration point j' to the center point o' in I ₂ (k) and the angle (d _r , θ _r ) between the straight line o'j' and the x-axis;

(5) Form the input training data V ^j = (d ₁ , θ ₁ , ..., d ₆ , θ ₆ , d′ ₁ , θ′ ₁ , ..., d′ ₆ , θ ′ ₆ , d _l , θ _l ) and the corresponding output training data V ^j = (Δd, Δθ) to form a training sample at time k Where Δd=d _r /d _l , Δθ=θ _r -θ _l ;

(6) Add j and j' as new matching points to the matching point set, and iterate the loop until all calibration points are traversed.

Compared with the prior art, the method of the present invention has the following prominent substantive features and significant advantages:

(1) In view of the shortcomings of the current face automatic calibration technology, which require high initial pose, many iterations, and slow calibration speed, the training data is used to obtain the spatial position relationship between the discrete feature points and the structured calibration points, and then according to the online The discrete feature points correspond to the initialization of the face calibration of the current frame;

(2) The kernel ridge regression method is used to obtain the mapping function between discrete feature points and structured calibration points, which achieves a good balance in regression accuracy and speed;

(3) By establishing the mapping relationship between discrete feature points and structured calibration points, the initialization of face calibration points can be obtained according to the correspondence of online discrete feature points, which can improve the speed and accuracy of final calibration;

The active appearance model initialization method provided by the present invention can greatly improve the positioning and tracking accuracy of the target shape, provide more comprehensive and accurate feature point information for the subsequent processing of face analysis, and achieve an ideal calibration effect. It has broad application prospects in civil and military fields such as intelligent video conferencing, film and television production, and public place safety monitoring.

Description of drawings

FIG. 1 is a schematic flow chart of the regression-based active appearance model initialization method of the present invention.

Figure 2 shows the tracking results of sample frame images in the FGNet Talking Face video.

Fig. 3 shows the tracking error comparison between the AAM tracking result based on the initialization method proposed by the present invention and the traditional AAM tracking result.

Detailed ways

The present invention will be further described below in conjunction with the specific illustration in FIG. 1 .

With reference to the flow chart in Fig. 1, realize the regression-based active appearance model initialization method of the present invention, the algorithm is first in the training phase, utilizes the discrete local feature point corresponding to the structured calibration point that Kernel Ridge Regression algorithm (Kernel Ridge Regression, KRR) establishes Then in the test phase, assuming that the key feature points of the first frame are known, the double-threshold method is used to obtain the accurate local feature points corresponding to the previous frame and the current frame image, and then the discrete local features obtained by training are used The mapping relationship between the point correspondence and the structured calibration point, and the calibration point positioning information of the current frame is extracted from the local feature correspondence. The specific implementation of each step is now described:

1. Select the training video, and use the kernel ridge regression algorithm to establish the mapping M _v of the spatial position between the scattered local feature points and the facial feature calibration points. The specific steps of the process are:

(1) Data initialization, let time k=0;

(2) Between the previous frame face image I ₁ (k) and the current face image I ₂ (k), scattered matching points are obtained by establishing a matching method of an equalized probability model, so that k=k+1;

(3) Acquire KRR training data from the previous frame face image and the current face image according to the scattered matching points Among them, m=66 is the number of calibration points; the specific implementation steps are:

(a) Let p be the current calibration point in the front face, o be the center of the current k matching points in the front face, and o' be the center of the current k matching points in the corresponding side face;

(b) Calculate the distance from the matching point i (the value of i is the number of matching points) to the center point o of the matching point and the angle between the straight line oi and the x-axis (d _i , θ _i ), and from the matching point The distance from i' to the center point o' of the matching point and the angle between the straight line o'i' and the x-axis (d' _i , θ' _i );

(c) Calculate the distance from the calibration point p to the center point o and the angle (d _l , θ _l ) between the straight line op and the x-axis in the frontal face;

(d) Calculate the distance from the calibration point p' to the center point o' in the side face and the angle (d _r , θ _r ) between the straight line o'p' and the x-axis;

(e) Form the input training data N _x =(d ₁ , θ ₁ , ..., d ₆ , θ ₆ , d′ ₁ , θ′ ₁ , ..., d′ ₆ , θ ' ₆ , d _l , θ _l ) and the corresponding output training data N _y =(Δd, Δθ), where Δd=d _r /d _l , Δθ=θ _r -θ _l ;

(f) Add p and p' as new matching points to the matching point set, and iterate the loop until all calibration points are traversed.

(4) If k<n, return to (1.2), otherwise get the total training data Where n=100 is the logarithm of training samples.

(5) For each face calibration point j, first according to the training data Calculate the kernel function matrix K _j , where k ₁ =1, 2...n, k ₂ ==1, 2...n, where σ=0.025; then create an identity matrix I of the same size as the matrix K _j , where I(k ₁ , k ₂ ) =1, k ₁ =1, 2,...n, k ₂ ==1, 2,...n; then calculate the kernel coefficient vector α _j , where where λ=0.5×10 ^-7 ; finally, according to the above calculation, the regression kernel function is obtained

(6) Obtain the mapping set M _v ={f ^j (V), j=1, 2, . . . , m}.

2. Use the CascadeAdaboost algorithm to detect faces, and normalize the face image to a size of 250*250;

3. Calculate the error between the reconstructed image of the previous frame and the current reconstructed image Where I ₁ and I ₂ are the face image of the previous frame and the current face image respectively, x is the pixel set under the mean shape _s0 , p is the deformation parameter from the mean shape _s0 to the current reconstructed shape s, W(x; p) is the set of pixels under the reconstruction shape s, and N is the number of pixels under the mean shape. When e>e ₀ , it means that the face image in the previous frame is quite different from the current face image, and go to step (3); otherwise, go to step (5), where e ₀ =5e-4 is the error threshold.

4. Extract the SIFT features of the previous frame face image I ₁ and the current frame face image I ₂ , use the feature matching method based on double thresholds to match, and obtain the matching pair C={(c _k , c′ _k ), k= 1, 2, ..., q _C }, where q _C is the number of matching pairs. The specific steps of the process are:

(1) Double threshold initialization: set the threshold initial value η ₁ =1.5, η ₂ =8; the number of cycles iter ₁ =0, iter ₂ =0; the number of cycles iter_max ₁ =10, iter_max ₂ =20, adjacent matching points number t=5;

(2) Extract SIFT feature from previous frame face image I ₁ and current face image I ₂ ;

(3) Utilize threshold η ₁ to do SIFT feature matching, and make iter ₁ =iter ₁ +1;

(4) If the number of matches is less than t, and iter ₁ <iter_max ₁ , set η ₁ =η ₁ +0.15, and return to (4.3); if the number of matches is greater than t, or iter ₁ >iter_max ₁ , then transfer to (4.5);

(5) Utilize threshold η ₂ to do SIFT feature matching, and make iter ₂ =iter ₂ +1;

(6) If the number of matches is less than 2, and iter ₂ <iter_max ₂ , set η ₂ =η ₂ +0.02, and return (4.5); if the number of matches is greater than 2 and less than 5, and iter ₂ <iter_max ₂ , Then set η ₂ =η ₂ -0.01, and return to (4.5); otherwise, turn to (4.7);

( ₇ ) Obtain the dense matching set B={(b _j , b′ _j ), j=1, 2, ..., q _B } obtained by using η 1, where q _B is the number of dense matching, and using η ₂ The obtained exact matching set A={(a _i , a' _i ), i=1, 2, ..., q _A }, where q _A is the number of dense matches;

(8) Calculate the distance ratio of the first and last two pairs of matching points in the exact matching set

(9) For each pair of matches (b _j , b′ _j ) in the dense matching set, first calculate then calculate and get Then calculate θ=atan((a _t (y)-b _j (y))/(a _t (x)-b _j (x))), θ'=atan((a' _t (y)-b' _j (y))/(a _t (x)-b _j (x))), and calculate sig _x = sign(a _t (y)-b _j (y)), sig _y = sign(a _t (x )-b _j (x)), sig' _x = sign(a' _t (y)-b' _j (y)), sig _y = sign(a' _t (x)-b' _j (x)). if then save the current match, otherwise delete the match. Finally, the final matching pair C={(c _k , c′ _k ), k=1, 2, . . . , q _C } is obtained, where q _C is the number of final matching.

5. Extract a space vector V={V _k , k=1, 2, . . . , n} from the face image I ₁ of the previous frame, where n is the number of face feature points.

6. According to the mapping M _v parameters obtained in step 1 and the matching points obtained in step (4), set up the discrete feature point space position vector V in the test stage and send it into the mapping M _v as input, and output the corresponding human face mark Fixed point, that is, to get the initial value of the active appearance model used for the current frame tracking.

The present invention uses each frame image in the FGNet Talking Face video as a test image to verify the regression-based active appearance model initialization method proposed by the present invention.

Figure 2 shows the tracking results of sample frame images in the FGNet Talking Face video.

Fig. 3 provides the tracking error comparison based on the AAM tracking result based on the regression-based active appearance model initialization method proposed by the present invention and the tracking error of the traditional AAM tracking result, the error formula is shown in formula (1), wherein the real face calibration point coordinates is (x ⁰ _i , y ⁰ _i ), the coordinates of the face calibration points obtained by the algorithm are ( _xi , y _i ), where i=1,...,N, N is the number of calibration points, and the algorithm N=66 in this paper.

$\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; MN</span>}}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\sqrt{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

where M is the number of frames tracked.

It can be seen from the figure that, compared with the traditional AAM initialization method, the initialization method proposed by the present invention can help AAM obtain more accurate tracking results.

Claims (4)

1. A regression-based active appearance model initialization method, characterized in that: when utilizing the active appearance model to carry out automatic tracking of face feature points, assuming the target position information of the first frame in the known video tracking, using the local feature correspondence method Get the scattered local point feature correspondence of the front and rear frame images. The kernel ridge regression algorithm is used to establish the spatial position mapping relationship between scattered local feature points and facial feature calibration points, so as to complete the initialization of the active appearance model. The specific implementation steps are as follows: (1) Select the training video, and use the kernel ridge regression algorithm to establish the mapping M _v of the spatial position between the scattered local feature points and the facial feature calibration points; (2) Use the Cascade Adaboost algorithm to detect faces, and normalize the face image to a size of 250*250; (3) Calculate the error between the reconstructed image of the previous frame and the current reconstructed image Where I ₁ and I ₂ are the face image of the previous frame and the current face image respectively, x is the pixel set under the mean shape _s0 , p is the deformation parameter from the mean shape _s0 to the current reconstructed shape s, W(x; p) is the set of pixels under the reconstruction shape s, and N is the number of pixels under the mean shape; when e>e ₀ , it shows that the face image of the previous frame is quite different from the current face image, and then go to step (3), Otherwise, go to step (5), where e ₀ =5e-4 is the error threshold; (4) Extract the SIFT features of the face image I ₁ of the previous frame and the face image I ₂ of the current frame, and use the feature matching method based on double thresholds to match, and obtain the matching pair C={(c _k , c′ _k ), k =1, 2, ..., q _C }, where q _C is the number of matching pairs; (5) Extract space vector V={V _k , k=1, 2,..., n} in the previous frame face image I ₁ , wherein n is the number of face feature points; (6) According to the mapping M _v parameters obtained in step (1) and the matching points obtained in step (4), the spatial position vector V of the discrete feature points in the test stage is established as input and sent to the mapping M _v , and the corresponding output The face calibration point is to obtain the initial value of the active appearance model used for the current frame tracking. 2. the regression-based active appearance model initialization method according to claim 1, wherein step (1) is carried out as follows: (1.1) Data initialization, let time k=0; (1.2) Between the face image I ₁ (k) of the previous frame and the current face image I ₂ (k), scattered matching points are obtained by establishing a matching method of an equalized probability model, so that k=k+1; (1.3) Acquire KRR training data from the face image of the previous frame and the current face image according to the scattered matching points $T_{V_j}\left(k\right)=\{(V_k^j,V_{k^{\&prime;}}^j),j=\mathrm{1,2},...,m\},$ Among them, m=66 is the number of calibration points; (1.4) If k<n, return to (1.2), otherwise get the total training data Among them, n=100 is the logarithm of training samples; (1.5) For each face calibration point j, first according to the training data Calculate the kernel function matrix K _j , where k ₁ =1, 2,...n, k ₂ ==1, 2,...n, where σ=0.025; then create an identity matrix I with the same size as the matrix K _j , where I(k ₁ , k ₂ )=1, k ₁ =1, 2,...n, k ₂ ==1, 2,...n; then calculate the kernel coefficient vector α _j , where where λ=0.5×10 ^-7 ; finally, according to the above calculation, the regression kernel function is obtained (1.6) Obtain the mapping set M _v ={fj(V), j=1, 2, . . . , m}. 3. the regression-based active appearance model initialization method according to claim 2, wherein step (1.3) is carried out as follows: (1.3.1) Suppose j is the current calibration point in I ₁ (k), o is the center of the current t adjacent matching points in I ₁ (k), and o′ is the current t matching points in I ₂ (k) corresponding to it match point center; (1.3.2) Calculate the distance from the matching point i (the value of i is the number of matching points) to the center point o of the matching point and the angle between the straight line oi and the x-axis (d _i , θ _i ), and from The distance from the matching point i' to the center point o' of the matching point and the angle between the straight line o'i' and the x-axis (d' _i , θ' _i ); (1.3.3) Calculate the distance from the calibration point j to the center point o in I ₁ (k) and the angle (d _l , θ _l ) between the straight line oj and the x-axis; (1.3.4) Calculate the distance from the calibration point j' to the center point o' in I ₂ (k) and the angle (d _r , θ _r ) between the straight line o'j' and the x-axis; (1.3.5) Form the input training data V ^j =(d ₁ , θ ₁ , . . . , d ₆ , θ ₆ , d′ _i , θ′ ₁ , . . . , d′ ₆ , θ′ ₆ , d _l , θ _l ) and the corresponding output training data V ^j = (Δd, Δθ) to form a training sample at time k Where Δd=d _r /d _l , Δθ=θ _r -θ _l ; (1.3.6) Add j and j' as new matching points to the matching point set, and iterate the loop until all calibration points are traversed. 4. the regression-based active appearance model initialization method according to claim 1, wherein step (4) is carried out as follows: (4.1) Double threshold initialization: set the threshold initial value η ₁ =1.5, η ₂ =8; the number of cycles iter ₁ =0, iter ₂ =0; the number of cycles iter_max ₁ =10, iter_max ₂ =20, adjacent matching points number t=5; (4.2) Extract SIFT features from the previous frame face image I ₁ and the current face image I ₂ ; (4.3) Utilize threshold η ₁ to do SIFT feature matching, and make iter ₁ =iter ₁ +1; (4.4) If the number of matches is less than t, and iter ₁ <iter_max ₁ , set η ₁ =η ₁ +0.15, and return to (4.3); if the number of matches is greater than t, or iter ₁ >iter_max ₁ , turn to (4.5); (4.5) Utilize threshold η ₂ to do SIFT feature matching, and make iter ₂ =iter ₂ +1; (4.6) If the number of matches is less than 2, and iter ₂ <iter_max ₂ , set η ₂ =η ₂ +0.02, and return to (4.5); if the number of matches is greater than 2 and less than 5, and iter ₂ <iter_max ₂ , Then set η ₂ =η ₂ -0.01, and return to (4.5); otherwise, turn to (4.7); ( _4.7 ) Obtain the dense matching set B={(b _j , b′ _j ), j=1, 2, ..., q _B } obtained by using η 1, where q _B is the number of dense matching, and using η ₂ The obtained exact matching set A={(a _i , a' _i ), i=1, 2, ..., q _A }, where q _A is the number of dense matches; (4.8) Calculate the distance ratio of the first and last two pairs of matching points in the exact matching set (4.9) For each pair of matches (b _j , b′ _j ) in the dense matching set, first calculate then calculate and get Then calculate θ=atan((a _t (y)-b _j (y))/(a _t (x)-b _j (x))), θ'=atan((a' _t (y)-b' _j (y))/(a _t (x)-b _j (x))), and calculate sig _x = sign(a _i (y)-b _j (y)), sig _y = sign(a _t (x )-b _j (x)), sig′ _x = sign(a′ _t (y)-b′ _j (y)), sig _y = sign(a′ _t (x)-b′ _j (x)); if Then save the current match, otherwise delete the match; finally get the final match pair C={(c _k , c′ _k ), k=1, 2, . . . , q _C }, where q _C is the number of final matches.