CN112528902B - Video monitoring dynamic face recognition method and device based on 3D face model - Google Patents

Abstract

The invention discloses a video monitoring dynamic face recognition method and a video monitoring dynamic face recognition device based on a 3D face model, wherein the method comprises the following steps: extracting two-dimensional features of a two-dimensional face image to be recognized, which are acquired by an acquisition end, converting the two-dimensional features into a three-dimensional face image, extracting three-dimensional features of the three-dimensional face image obtained by conversion, and connecting the three-dimensional features in series to obtain a first fusion feature containing two-dimensional information and three-dimensional information; simultaneously, extracting three-dimensional information of a three-dimensional face model prestored in the identification terminal, projecting the prestored three-dimensional face model to a two-dimensional projection image, extracting two-dimensional features of the projection image, connecting in series to obtain a second fusion feature containing the two-dimensional information and the three-dimensional information, and finally performing face identification by using the two fusion features; the corresponding fusion characteristics fully fuse the three-dimensional shape information on the basis of the two-dimensional texture information, the problem that the identification cannot be successfully identified in a complex environment only by using the two-dimensional texture information is solved, the identification accuracy is effectively improved, and the robustness of an identification algorithm is ensured.

Description

A method and device for dynamic face recognition in video surveillance based on 3D face model

technical field

The invention relates to the technical field of computer vision and pattern recognition, in particular to a dynamic face recognition method and device for video surveillance based on a 3D face model.

Background technique

Face recognition technology has become a hot spot and a bright spot in the research, development and application of a new generation of artificial intelligence. Thanks to the rapid development of new-generation artificial intelligence technologies such as big data and deep learning, face recognition technology based on two-dimensional (2D) pictures has achieved great success in many application fields with controllable environment and user cooperation, such as security inspection and finance. Important applications have produced huge social and economic benefits. However, for a wider range of applications in dynamic scenarios where the environment is uncontrollable and users are not cooperative, such as video surveillance dynamic face recognition, the performance of the existing face recognition technology is still far from meeting the application requirements. Face recognition technology based on three-dimensional (3D) face model is one of the future development trends. 3D face model has more abundant information such as three-dimensional shape than 2D face image, which can improve the environment such as uncontrollable environment and user non-cooperation. Recognition performance in dynamic situations such as large poses and multiple illumination changes. However, both the registration terminal and the recognition terminal use 3D sensors to capture 3D faces, and it cannot be achieved in a short time to transform all 2D cameras in the current society into 3D sensors. A practical solution is that the registration terminal collects a 3D face model, and the recognition terminal collects one or more 2D face pictures for identification, that is, a related technology of using a three-dimensional face model to identify two-dimensional face pictures.

At present, the technologies involved in using 3D face models to recognize 2D face images are relatively scarce, and most of them are still at the stage of using 3D face models to recognize 3D faces, rather than using 3D face models to recognize 2D face images. The Chinese invention patent application with the application publication number CN108427871A discloses a 3D face fast identity authentication method and device, which rotates the three-dimensional face model to the same posture as the two-dimensional image to be recognized and projects it to the two-dimensional image, and then projects the projection onto the two-dimensional image. The latter two-dimensional image is compared and identified with the two-dimensional image to be identified. The Chinese invention patent application with the application publication number CN109858433A discloses a method and device for recognizing a two-dimensional face picture based on a three-dimensional face model, which projects the three-dimensional face model and a single two-dimensional picture to multiple poses through a certain criterion Then carry out matching identification respectively. However, the above methods only project the 3D face model to the 2D image, use the 3D information to align the 2D images, and finally use the aligned 2D texture information to perform face recognition, but there is no feature comparison process. The 3D shape information contained in the 3D face is directly used for face recognition.

SUMMARY OF THE INVENTION

The object of the present invention is to overcome the defect of not using three-dimensional shape information for face recognition existing in the existing technology of using three-dimensional face model to identify two-dimensional face images, and to provide a video monitoring dynamic face based on 3D face model. The identification method and device, on the basis of extracting the two-dimensional information of the two-dimensional image to be identified and the two-dimensional projection information corresponding to the three-dimensional information of the pre-stored three-dimensional face model, convert the two-dimensional image to be identified collected by the collection end into 3D face model, and perform UV conversion on the 3D face model to obtain 3D information. At the same time, UV conversion is used to extract the 3D information of the 3D face model pre-stored at the recognition terminal, and then the fusion features of 2D information and 3D information are used for feature comparison. , face recognition; face recognition by combining two-dimensional texture information and three-dimensional shape information, to improve the problem that only using two-dimensional projection images (only using two-dimensional texture information) for recognition cannot be successfully recognized in complex environments, Effectively improve the accuracy of recognition.

In order to achieve the above-mentioned purpose of the invention, the present invention provides the following technical solutions:

A dynamic face recognition method for video surveillance based on a 3D face model, comprising:

A. Extract N two-dimensional face images to be identified; use the first feature extractor to perform feature extraction on N described two-dimensional face images to be identified to obtain a first two-dimensional feature vector; wherein, N is an integer and N ≥1;

And, converting the N two-dimensional face images to be identified into a three-dimensional face model, carrying out UV expansion on the obtained three-dimensional face model, obtaining a first UV map, and utilizing a second feature extractor to characterize the first UV map. extracting to obtain a first three-dimensional feature vector; connecting the first two-dimensional feature vector and the first three-dimensional feature vector to obtain a first fusion feature vector;

B. Project the pre-stored three-dimensional face model to the corresponding perspective of the N two-dimensional face images to be recognized, to obtain N two-dimensional projected face images, and use the first feature extractor to perform the N two-dimensional projected face images. Perform feature extraction on the face image to obtain a second two-dimensional feature vector;

and, performing UV expansion on the pre-stored three-dimensional face model to obtain a second UV image, and using a second feature extractor to perform feature extraction on the second UV image to obtain a second three-dimensional feature vector; the second two-dimensional feature vector is connected in series and the second three-dimensional feature vector to obtain the second fusion feature vector;

C. Perform feature comparison between the first fused 3D feature vector and the second fused 3D feature vector to obtain a face recognition result.

Preferably, in the above-mentioned dynamic face recognition method for video surveillance based on a 3D face model, the extracting N two-dimensional face images to be recognized includes: tracking and recognizing the faces in the video surveillance, obtaining a face video stream, based on The preset screening condition selects N pieces of the to-be-recognized two-dimensional face images from the face video stream.

Preferably, in the above-mentioned dynamic face recognition method for video surveillance based on 3D face model, when N>1, the extracted N feature vectors of the two-dimensional face images to be recognized are connected in series to obtain the the first two-dimensional feature vector;

When N>1, the second two-dimensional feature vector is obtained by concatenating the extracted N feature vectors of the two-dimensional projected face images.

Preferably, in the above-mentioned dynamic face recognition method for video surveillance based on a 3D face model, the step C further comprises: using a multi-layer perceptron to characterize the first fusion feature vector and the second fusion feature vector respectively Transformation and dimensionality reduction processing are performed to compare the features of the second fusion feature vector after dimensionality reduction processing with the first fusion feature vector to obtain a face recognition result.

Preferably, in the above-mentioned dynamic face recognition method for video surveillance based on a 3D face model, feature comparison is performed by calculating the cosine similarity or Euclidean distance between the second fusion feature vector after dimensionality reduction processing and the first fusion feature vector.

Preferably, in the above-mentioned dynamic face recognition method for video surveillance based on a 3D face model, the UV conversion includes: converting the shape information of the 3D face model into a UV position map and a UV normal vector map.

Preferably, in the above-mentioned dynamic face recognition method for video surveillance based on 3D face model, the value of the pixel on the UV position map corresponds to the coordinates of each 3D point on the 3D face model, and the UV normal vector map The value of the upper pixel corresponds to the normal vector of each 3D point on the 3D face model.

Preferably, in the above-mentioned dynamic face recognition method for video surveillance based on 3D face model, the first feature extractor and the second feature extractor are convolutional neural networks VGG-16, ResNet-50 or ResNet-101 one of the.

In a further embodiment of the present invention, there is also provided a dynamic face recognition device for video surveillance based on a 3D face model, comprising at least one processor and a memory communicatively connected to the at least one processor; the memory stores There are instructions executable by the at least one processor, and the instructions are executed by the at least one processor, so that the at least one processor can execute the above-mentioned dynamic face recognition method for video surveillance based on a 3D face model .

Compared with the prior art, the beneficial effects of the present invention:

The face recognition method provided by the present invention extracts the two-dimensional features of the two-dimensional image to be recognized collected by the collection end, converts the to-be-recognized image into a three-dimensional face model, and extracts the three-dimensional features of the converted three-dimensional face model, The two are connected in series to obtain the first fusion feature containing two-dimensional information and three-dimensional information; at the same time, the three-dimensional information of the three-dimensional face model pre-stored at the recognition end is extracted, and the pre-stored three-dimensional face model is projected to the two-dimensional projection image, and the projection is extracted. The two-dimensional features of the image are connected in series to obtain the second fusion feature containing the two-dimensional information and the three-dimensional information. Finally, the features that integrate the two-dimensional information and the three-dimensional information can be directly used for face recognition; On the basis of 3D texture information, the 3D shape information is fully integrated, which improves the problem that only 2D projection images (using only 2D texture information) can not be successfully recognized in complex environments, and effectively improves the accuracy of recognition and guarantees Robustness of the recognition algorithm.

In the scene of using multiple video surveillance face images of the same person for face tracking and identification, compared with the identification method using only two-dimensional texture features, the fusion feature calculated by the present invention has stronger robustness, and the human Compared with the existing recognition methods, the face recognition effect is better, and the applicable scenarios are more versatile.

Description of drawings:

FIG. 1 is a flowchart of a method for dynamic face recognition in video surveillance based on a 3D face model according to an exemplary embodiment 1 of the present invention.

FIG. 2 is a schematic block diagram of a method for dynamic face recognition in video surveillance based on a 3D face model according to an exemplary embodiment 2 of the present invention.

FIG. 3 is a structural diagram of a dynamic face recognition device for video surveillance based on a 3D face model according to an exemplary embodiment of the present invention.

Detailed ways

The present invention will be further described in detail below in conjunction with test examples and specific embodiments. However, it should not be construed that the scope of the above-mentioned subject matter of the present invention is limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

Example 1

1 shows a dynamic face recognition method for video surveillance based on a 3D face model according to an exemplary embodiment of the present invention, including:

A. Extract N two-dimensional face images to be identified; use the first feature extractor to perform feature extraction on N described two-dimensional face images to be identified to obtain a first two-dimensional feature vector; wherein, N is an integer and N ≥1;

And, converting the N two-dimensional face images to be identified into a three-dimensional face model, carrying out UV expansion on the obtained three-dimensional face model, obtaining a first UV map, and utilizing a second feature extractor to characterize the first UV map. extracting to obtain a first three-dimensional feature vector; connecting the first two-dimensional feature vector and the first three-dimensional feature vector to obtain a first fusion feature vector;

B. Project the pre-stored three-dimensional face model to the corresponding perspective of the N two-dimensional face images to be recognized, to obtain N two-dimensional projected face images, and use the first feature extractor to perform the N two-dimensional projected face images. Perform feature extraction on the face image to obtain a second two-dimensional feature vector;

and, performing UV expansion on the pre-stored three-dimensional face model to obtain a second UV image, and using a second feature extractor to perform feature extraction on the second UV image to obtain a second three-dimensional feature vector; the second two-dimensional feature vector is connected in series and the second three-dimensional feature vector to obtain the second fusion feature vector;

C. Perform feature comparison between the first fused 3D feature vector and the second fused 3D feature vector to obtain a face recognition result.

Specifically, the method of the present invention for performing face recognition using two-dimensional and three-dimensional fusion features will be described by selecting two-dimensional face images from three perspectives to be recognized. As shown in Figure 2, the method includes:

S1: Track faces in video surveillance, and select a specific number (N, N is an integer ≥ 1) representative two-dimensional face images for the person to be identified for subsequent identification.

Specifically, face detection and tracking are methods well known to those skilled in the art, and there are many such technical methods, which can track the faces in the surveillance video to obtain a continuous sequence of each face. After obtaining a continuous sequence of faces to be recognized, this embodiment automatically selects a fixed specific number (N) of representative two-dimensional face images after comprehensive analysis of the resolution, image quality, attitude angle, etc. of the face through an algorithm for subsequent identification. In this embodiment, the selected representative two-dimensional face image represents the characteristics of the face under different angles and postures as much as possible. In the present invention, N is an integer greater than or equal to 1, N=1 means single view, N>1 means multi-view, generally only a multi-view two-dimensional image can reconstruct a better three-dimensional face model, so N>1 is preferred, However, if the quality of the selected single-view image can meet the requirements of three-dimensional reconstruction, then a single-view image can also be selected, that is, N=1. In this embodiment, N=3 is used for description.

S2: using a neural network to extract the two-dimensional feature vectors of the N two-dimensional face images, and then performing feature series to obtain a first two-dimensional feature vector;

Specifically, for each face image to be recognized (3 images in total), a first feature extractor, that is, a convolutional neural network F, is used to extract a face feature vector, and then the separately extracted face feature vectors are connected in series. Likewise, the vector finally obtained in this step is 256*3=768 dimensions. F is a trained neural network for face feature extraction, which can be a common network structure for face feature extraction such as VGG-16, ResNet-50, and ResNet-101.

S3: A method of generating a 3D face model by using a single or multiple 2D images, generating a 3D face model from a single-view or multi-view 2D face image, and performing UV conversion on the generated 3D face model to convert to UV Position map and UV normal map.

Specifically, generating a three-dimensional face model from a single or multiple two-dimensional face images is a method well known to those skilled in the art. There are many such methods, such as "3D Face Reconstruction Using a Single orMultiple Views", "Examplar coherent 3D face reconstruction from forensicmugshot database", "Automated 3D Face Reconstruction from Multiple Images using Quality Measures", "Fast,Approximate 3D Face Reconstruction fromMultiple" Views" and other proposed methods can be used. For the generated 3D face model, the 3D face model is converted into a UV position map by the method described in the document "Deep 3D Facial Landmark Localization on position maps", and each point on the UV position map is converted into a UV position map. The RGB color value represents the normalized (X, Y, Z) coordinates of the 3D point on the corresponding 3D face model, and the 3D point and pixel coordinates are in a one-to-one correspondence. In the same way, the normal vector (NX, NY, NZ) of each 3D point is calculated and converted to a UV normal vector map in a similar way. The generated 3D face model is mapped to the UV coordinate system (or called UV unwrapping of the 3D face model of the face), and converted into a first UV map, including: a first UV position map and a first UV normal vector map.

S4: The UV position map and the UV normal vector map obtained from S3 are connected in series on the channel to obtain a first UV map, and a neural network is used to extract the first three-dimensional feature vector of the first UV map.

Specifically, the second feature extractor is used to extract the first three-dimensional feature vector in the first UV map. The second feature extractor used is a neural network G, G is a trained neural network for extracting three-dimensional face feature vectors from UV images, which can be VGG-16, ResNet-50, ResNet-101, etc. The network structure of face feature extraction, the extraction vector in this embodiment is 512 dimensions. Among them, the first feature extractor and the second feature extractor are consistent in the network architecture, and they are one of VGG-16, ResNet-50 or ResNet-101. Finally, the first two-dimensional feature vector obtained by S2 and the first three-dimensional feature vector obtained by S4 are connected in series on the channel to obtain the first fusion feature vector.

At the same time, feature extraction is performed on the 3D face model pre-stored at the recognition end, including: S5, projecting the pre-stored 3D face model to the corresponding poses of the N face images, and using a neural network to extract the features of each projected image vector, and then further feature concatenation to obtain the second two-dimensional feature vector

(这里采用文献《Fine-Grained Head Pose EstimationWithout Keypoints》提出的方法估计人脸的姿态角)，其中ω表示偏航角(yaw)，θ表示翻滚角(roll)，

表示俯仰角(pitch)。本实施例中N＝3，则用该方法估计每一张二维人脸图像的姿态角。在获得每张图片的姿态角后，将三维人脸模型分别投影至所述每张待识别人脸图像相应的姿态角，三维人脸模型投影��二维图像为本领域技术人员熟知的方法。��后，每一张投影������分别通过同一卷积神经网络F提取人脸特征向量，本实施例中所提取向量维256维。因本实施例中N＝3，则将分别提取的人脸特征向量进行串联，最终得到的第二二维特征向量为256*3＝768维。Specifically, the N face images to be recognized are two-dimensional images, and the recognition library contains three-dimensional face models. The face recognition problem in this embodiment is to answer the image to be recognized and a three-dimensional face model belonging to the library. Whether it is matched or not, and get the matching score, that is, use the three-dimensional face model to identify single or multiple two-dimensional face images. First, estimate the pose angle of the face in the face image to be recognized

(Here, the method proposed in the document "Fine-Grained Head Pose EstimationWithout Keypoints" is used to estimate the pose angle of the face), where ω represents the yaw angle (yaw), θ represents the roll angle (roll),

Indicates the pitch angle (pitch). In this embodiment, N=3, and this method is used to estimate the pose angle of each two-dimensional face image. After obtaining the attitude angle of each picture, the three-dimensional face model is projected to the corresponding attitude angle of each face image to be recognized, and the method of projecting the three-dimensional face model to the two-dimensional image is well known to those skilled in the art. After that, each projection image is respectively extracted through the same convolutional neural network F to extract the face feature vector, and the dimension of the extracted vector is 256 in this embodiment. Since N=3 in this embodiment, the face feature vectors extracted respectively are connected in series, and the second two-dimensional feature vector finally obtained is 256*3=768 dimensions.

S6: Convert the shape information of the pre-stored 3D face model into a UV position map and a UV normal vector map. The value of the pixel on the UV position map corresponds to the (X, Y, Z) coordinates of the 3D point on the 3D face model, and the UV The value of the pixel on the normal vector map corresponds to the normal vector (NX, NY, NZ) of the 3D point on the 3D face model.

Specifically, the three-dimensional face model is converted into a UV position map by the method described in the document "Deep 3D Facial Landmark Localization onposition maps", and the RGB color value of each point on the UV position map represents the corresponding three-dimensional Normalized (X, Y, Z) coordinates of 3D points on the face model. In the same way, the normal vector (NX, NY, NZ) of each 3D point is calculated and converted to a UV normal vector map in a similar way. The 3D face model of the face is mapped to the UV coordinate system (or called UV unwrapping of the 3D face model of the face).

S7: The UV position map and the UV normal vector map obtained from S6 are concatenated on the channel, and then the three-dimensional feature vector is extracted by the neural network.

Specifically, the aforementioned neural network G is used to extract the second three-dimensional face feature vector from the UV position map and the UV normal vector map in series, and the vector dimension is 512 in this embodiment. The second two-dimensional feature vector and the second three-dimensional feature vector are concatenated on the channel to obtain the second fusion feature vector.

S8: Perform feature transformation and dimensionality reduction on the two concatenated features through a shared multi-layer perceptron (MLP) respectively, and finally obtain a first feature vector and a second feature vector for matching and identification.

Specifically, the 768-dimensional two-dimensional feature obtained in S2 and the 512-dimensional three-dimensional feature obtained in S4 are connected in series to obtain a 1280-dimensional feature A. At the same time, the 768-dimensional two-dimensional feature obtained in S5 and the 512-dimensional three-dimensional feature obtained in S7 are connected in series to obtain a 1280-dimensional feature B. Use the trained three-layer perceptron M (the three layers respectively contain 1024, 1024, 256 neurons) to convert the 1280-dimensional features into the final 256-dimensional features; that is, input A and B into M respectively, and get them for matching and recognition. The first and second eigenvectors of .

S9: Compare the first feature vector obtained according to the three-dimensional face model with the second feature vector obtained according to N representative two-dimensional face images, and obtain a matching score for face verification or identification.

The matching score can be obtained by calculating the cosine similarity or Euclidean distance between the two 256-dimensional vectors processed in s8.

In a further embodiment of the present invention, steps S1-S4 and S5-S7 may be performed in sequence, or may be extracted simultaneously, and the sequence of operations is not specific.

In the above embodiment, by converting the three-dimensional information into the UV position map and the UV normal vector for identification, the three-dimensional shape information can be extracted, and then fused with the extracted two-dimensional information to obtain the fusion feature information, and the fusion information can be used for human Face Recognition Improvement Only use projected images and single video surveillance face images to recognize the problems of unrecognized or incorrect recognition that may occur in complex environments, and improve the accuracy of the system. Moreover, the identification method provided by the present invention has stronger robustness to the scene of face video stream tracking identification, and has a wider scope of application and applicable scene compared with the identification method using only two-dimensional texture information. It is more versatile and has better recognition effect.

Example 2

FIG. 3 shows an apparatus for recognizing a two-dimensional face picture based on a three-dimensional face model according to an exemplary embodiment of the present invention, that is, an electronic device 310 (eg, a computer server with a program execution function), which includes at least one processor 311, a power supply 314, and a memory 312 and an input-output interface 313 communicatively connected to the at least one processor 311; the memory 312 stores instructions executable by the at least one processor 311, the instructions being executed by the at least one processor 311 At least one processor 311 executes, so that the at least one processor 311 can execute the method disclosed in any of the foregoing embodiments; the input and output interface 313 may include a display, a keyboard, a mouse, and a USB interface for input and output Data; power supply 314 is used to provide power to electronic device 310 .

Those skilled in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by program instructions related to hardware, the aforementioned program can be stored in a computer-readable storage medium, and when the program is executed, the execution includes the above The steps of the method embodiment; and the aforementioned storage medium includes: a removable storage device, a read only memory (Read Only Memory, ROM), a magnetic disk or an optical disk and other media that can store program codes.

When the above-mentioned integrated units of the present invention are implemented in the form of software functional units and sold or used as independent products, they may also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present invention may be embodied in the form of software products in essence or the parts that make contributions to the prior art. The computer software products are stored in a storage medium and include several instructions for A computer device (which may be a personal computer, a server, or a network device, etc.) is caused to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a removable storage device, a ROM, a magnetic disk, or an optical disk.

The above description is only a detailed description of the specific embodiments of the present invention, rather than a limitation of the present invention. Various substitutions, modifications and improvements made by those skilled in the relevant technical field without departing from the principle and scope of the present invention should be included within the protection scope of the present invention.

Claims (9)

1. a video surveillance dynamic face recognition method based on 3D face model, is characterized in that, comprises: A. Extract N two-dimensional face images to be identified; use the first feature extractor to perform feature extraction on N described two-dimensional face images to be identified to obtain a first two-dimensional feature vector; wherein, N is an integer and N ≥1; And, converting the N two-dimensional face images to be identified into a three-dimensional face model, carrying out UV expansion on the obtained three-dimensional face model, obtaining a first UV map, and utilizing a second feature extractor to characterize the first UV map. extracting to obtain a first three-dimensional feature vector; connecting the first two-dimensional feature vector and the first three-dimensional feature vector to obtain a first fusion feature vector; B. Project the pre-stored three-dimensional face model to the corresponding perspective of the N two-dimensional face images to be recognized, to obtain N two-dimensional projected face images, and use the first feature extractor to perform the N two-dimensional projected face images. Perform feature extraction on the face image to obtain a second two-dimensional feature vector; and, performing UV expansion on the pre-stored three-dimensional face model to obtain a second UV image, and using a second feature extractor to perform feature extraction on the second UV image to obtain a second three-dimensional feature vector; the second two-dimensional feature vector is connected in series and the second three-dimensional feature vector to obtain the second fusion feature vector; C. Perform feature comparison between the first fusion feature vector and the second fusion feature vector to obtain a face recognition result. 2. The method according to claim 1, wherein the extracting N two-dimensional face images to be identified comprises: tracking and identifying faces in video surveillance, obtaining face video streams, and screening based on preset Condition N pieces of the to-be-recognized two-dimensional face images are selected from the face video stream. 3. The method according to claim 1, wherein, when N>1, the N feature vectors of the extracted N two-dimensional face images to be identified are connected in series to obtain the first two-dimensional feature vector; when N>1, the second two-dimensional feature vector is obtained by concatenating the N feature vectors of the extracted N two-dimensional projected face images. 4. The method according to claim 1, wherein the step C further comprises: using a multilayer perceptron to perform feature transformation and dimension reduction on the first fusion feature vector and the second fusion feature vector respectively processing, to perform feature comparison between the second fusion feature vector after dimensionality reduction processing and the first fusion feature vector, to obtain a face recognition result. 5 . The method according to claim 4 , wherein the feature comparison is performed by calculating the cosine similarity or Euclidean distance of the second fusion feature vector after dimensionality reduction processing and the first fusion feature vector. 6 . 6. The method according to any one of claims 1-5, wherein the UV unwrapping comprises: converting the shape information of the three-dimensional face model into a UV position map and a UV normal vector map. 7. The method according to claim 6, wherein the value of the pixel on the UV position map is in one-to-one correspondence with the coordinates of each 3D point on the 3D face model, and the value of the pixel on the UV normal vector map is in one-to-one correspondence. One-to-one correspondence with the normal vector of each three-dimensional point on the three-dimensional face model. 8. The method of claim 1, wherein the first feature extractor and the second feature extractor are one of convolutional neural network VGG-16, ResNet-50 or ResNet-101 . 9. A video surveillance dynamic face recognition device based on a 3D face model, characterized in that it comprises at least one processor, and a memory connected in communication with the at least one processor; Instructions executed by at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 8.

Images