A three-dimensional face recognition method for underground coal mine

Abstract

The present application relates to a kind of three-dimensional face recognition methods for coal mine underground, belong to multidimensional stereoscopic face recognition technical field.The method is converted into three-dimensional point under camera coordinate system by two-dimensional face image, and rotation matrix is combined to carry out rotation operation, generates the rotated face side face data including multiple angles, provides rich data basis for three-dimensional reconstruction.PointNet++ and its extended structure are used in conjunction with auto-encoder framework to establish three-dimensional reconstruction model, can extract compact feature representation from three-dimensional point cloud, and utilize these features to reconstruct three-dimensional facial image with details and geometric structure, three-dimensional facial image is divided into several recognition regions, and feature recognition matching is carried out respectively;Identification, only need to filter clear area to carry out image reconstruction and feature matching, can directly give face recognition result.The present application improves the accuracy and fidelity of reconstruction.

Description

Technical Field

[0001] This invention belongs to the field of multi-dimensional facial recognition technology, and relates to a three-dimensional facial recognition method for use in underground coal mines. Background Technology

[0002] In the field of coal mine safety monitoring, the integration and application of information technology has become an important means to improve mine safety management and prevent accidents. With the increasing complexity and variability of the mine operating environment, traditional safety monitoring methods are no longer sufficient to meet the demands of modern coal mines for efficient and precise safety management. In recent years, biometric identification technology, especially facial recognition technology, has gradually emerged in coal mine safety monitoring due to its unique non-contact nature, high convenience, and user-friendliness. However, applying facial recognition technology to the special environment of coal mines still faces a series of technical challenges and difficulties.

[0003] Existing facial recognition technologies are mainly based on two-dimensional image processing and basic feature extraction methods, such as edge detection and feature point localization. Under ideal conditions, such as sufficient lighting, a direct angle, and natural facial expressions, these technologies can achieve good recognition results. However, the coal mine working environment is complex and variable, with poor lighting conditions, high dust concentrations, and workers' rich and varied facial expressions. These factors greatly affect the accuracy and robustness of facial recognition.

[0004] Firstly, regarding changes in lighting, insufficient or excessive lighting in coal mines can degrade facial image quality, thus affecting recognition accuracy. Existing 2D image processing technologies are often quite sensitive to changes in lighting, making it difficult to maintain stable recognition performance under complex lighting conditions.

[0005] Secondly, current facial recognition technology falls short when handling changes in angle and facial expression. Coal miners' head posture and facial expressions constantly change during operations, requiring the recognition system to have high adaptability to angles and expressions. However, current technology often experiences a significant drop in accuracy when dealing with large-angle turns or complex expressions.

[0006] Furthermore, occlusion is a significant challenge in facial recognition in coal mines. Due to the unique working environment, coal miners may wear protective gear such as safety helmets and masks, which can severely impact the integrity of facial images and the accuracy of feature extraction. Existing technologies often fail to accurately identify the occluded facial features when dealing with occlusion, leading to recognition failures.

[0007] With the continuous increase in coal mine safety monitoring data, existing facial recognition algorithms are facing severe challenges in terms of processing speed and real-time performance. In the context of big data, how to quickly and accurately process and analyze massive amounts of facial image data has become an urgent problem to be solved in the field of coal mine safety monitoring. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a three-dimensional face recognition method for underground coal mines, solving the following problems: (1) Existing two-dimensional image processing technologies are often sensitive to changes in illumination and are difficult to maintain stable recognition performance under complex lighting conditions; (2) Current technologies often experience a significant drop in recognition accuracy when dealing with large-angle deflections or complex expressions; (3) Existing technologies often fail to accurately identify the facial features of occluded parts when dealing with occlusion problems, leading to recognition failure. The method of this invention can improve the accuracy and consistency of face recognition.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A three-dimensional face recognition method for use in underground coal mines specifically includes the following steps:

[0011] S1: Use a high-resolution camera to capture diverse and wide-coverage original face images under different lighting and angles, and preprocess them, including filtering and noise reduction, grayscale normalization, etc., to output clean data images; then use the ICP registration method to register and fuse the same face images captured under different conditions to obtain a second face image, thereby improving the accuracy of subsequent processing.

[0012] S2: Divide the second face image into training set, validation set and test set according to the proportion, and use them for model learning, optimization and performance evaluation respectively; at the same time, generate feature point cloud data for building three-dimensional facial images from the second face image, including rotating face profile data from multiple angles.

[0013] S3: Establish a 3D reconstruction model, using PointNet++ and its extended structure combined with an autoencoder framework to extract features from 3D point clouds and reconstruct 3D facial images; through training, the model learns to extract useful information from point clouds and generate realistic 3D facial images.

[0014] S4: The trained 3D reconstruction model is used to identify the target faces of underground coal mine workers. The 3D facial image is divided into several recognition regions, and feature recognition and matching are performed separately. In the information database, the facial information of each person is stored separately according to the recognition region, with identity information. During recognition, only clear regions need to be selected for image reconstruction and feature matching, and the face recognition result can be given directly.

[0015] Furthermore, in step S1, the original face image is preprocessed, specifically including: filtering and denoising, grayscale normalization, edge detection or size normalization.

[0016] Further, in step S1, the ICP registration method involves geometrically transforming the same face images obtained at different times, in different image blocks, or from different shooting angles to ensure that the corresponding image points are completely superimposed in position and orientation; the specific calculation steps are as follows:

[0017] 1) Initialization: Among N facial images of the same face taken at different times or from different angles, a reference image I0 is selected as the baseline, and the feature point set P0 is extracted from this image; for the other N-1 images I1, I2, ..., I... N-1 Extract the corresponding feature point sets P1, P2, ..., P respectively. N-1 ;

[0018] 2) Registration process: For each pair of feature point sets P0 and P... i Perform the following steps:

[0019] 2.1) Find the nearest neighbor: For each point p0 in P0, find the nearest neighbor. i The nearest point p in the middle i It can be measured using Euclidean distance:

[0020] p i =arg min||p0-q||,q∈P i

[0021] 2.2) Calculate the transformation matrix: Based on the found nearest neighbor pair (p0, p... i Calculate the optimal rigid body transformation X i Rigid body transformation X i It typically includes a rotation matrix R. i Translation vector t i This can be solved using the least squares method:

[0022]

[0023] Where t is the translation vector; R is the rotation matrix. The rotation matrix is ​​calculated by first centering each point set, then calculating its covariance matrix, performing singular value decomposition (SVD) on the covariance matrix, and obtaining the rotation matrix from the SVD decomposition.

[0024] 2.3) Applying the transformation: Applying the calculated transformation R i and t i Applied to feature point set P i This yields a new set of points:

[0025] P i ′=R i ·P i +t i

[0026] 3) Iterative update: Repeat step 2) until the feature point set P is reached. i The alignment error with P0 is less than a certain threshold or the maximum number of iterations is reached.

[0027] By following the steps above, feature points in multiple images of the same person collected at different times and angles can be registered and fused, aligning key features such as eyes, nose, and mouth in different images to obtain a second face image. This improves the accuracy and consistency of subsequent processing steps.

[0028] Further, in step S2, feature point cloud data for building a 3D facial image is generated from the second face image. Specifically, this includes: knowing that the pixel value of each pixel in the second face image is I(x,y) and the initial depth value of each pixel is z(x,y), a virtual camera is set based on the second face image, with its intrinsic parameter matrix being K and its focal length relative to the second face image being f. x ,f y The center point of the image under the camera is c. x ,c y Then, each pixel (x, y) in the second face image is converted into a three-dimensional point (X, Y) in the camera coordinate system using the following formula. c ,Y c Z c ):

[0029]

[0030] Z c =z(x,y)

[0031] Then, rotate each point using a rotation matrix R; assuming the rotation matrix R contains three rotation angles θ. x ,θ y ,θ z Then the rotation matrix can be expressed as:

[0032] R = R z (θ z )·R y (θ y )·R x (θ x )

[0033] Among them, R x ,R y ,R z These are rotation matrices around the x, y, and z axes, respectively, and can be represented as:

[0034]

[0035] The rotation matrix R can be used to rotate each point in the camera coordinate system:

[0036]

[0037] Among them, (X) w ,Y w Z w () is a point in the world coordinate system.

[0038] Specifically, step S2 generates rotated face profile data from four angles, and selects 30 different face data with different poses (10°, 30°, and 45°) as a test set, and calculates the comparison error between the test set and the reference pose face data.

[0039] Furthermore, in step S3, the established 3D reconstruction model specifically includes:

[0040] The encoder part is mainly composed of PointNet++, which is responsible for extracting features from the input 3D point cloud. This feature extraction process is divided into multiple layers, and each layer captures the geometric structure information in the point cloud through local feature aggregation. Specifically, each point interacts with its neighboring points to calculate the local feature representation of that point. These local features are aggregated at multiple layers to finally obtain the global feature representation of the entire point cloud.

[0041] The decoder is responsible for converting the feature representations extracted by the encoder into a 3D mesh model or voxel model. This conversion process is usually achieved through a series of deconvolution operations or upsampling operations, which gradually restore the feature representations to points or voxels in 3D space. During the decoding process, the model uses the features extracted by the encoder to generate a 3D facial image with details and geometric structure.

[0042] Combining the encoder and decoder forms an autoencoder framework. In this framework, the model first extracts features from the input 3D point cloud through the encoder, and then restores these features into a 3D mesh model or voxel model through the decoder.

[0043] Furthermore, in step S3, training the 3D reconstruction model specifically includes: by training the model, it can learn how to effectively extract useful information from the point cloud and use this information to generate realistic 3D facial images.

[0044] The training process involves a large amount of 3D point cloud data and corresponding 3D mesh models or voxel models. This data is typically acquired through a 3D scanner or other acquisition devices. During training, the model attempts to minimize reconstruction error, aiming to generate a 3D model that is as close as possible to the real 3D model. Through continuous iteration and optimization, the model can gradually learn how to accurately reconstruct 3D facial images from point clouds.

[0045] Further, in step S4, the recognition area includes at least an eye recognition area, a nose recognition area, a mouth recognition area, an eye-nose recognition area, a mouth-nose recognition area, an ear-nose recognition area, and a face recognition area; the eye recognition area includes a left eye recognition sub-region and a right eye recognition sub-region, which respectively recognize and match features of the left and right eyes; the nose recognition area recognizes and matches nose features; the mouth recognition area recognizes and matches mouth features; the eye-nose recognition area refers to the facial area between the eyes and the nose, which includes a left eye nose recognition area and a right eye nose recognition area, and mainly judges the obvious features of the facial area between the eyes and the nose, as well as the nose. The facial recognition region includes the distance features between the nose and eyes; the mouth and nose recognition region refers to the facial area between the mouth and nose, which mainly judges the obvious features of the facial area between the nose and mouth, as well as the distance features between the nose and mouth and the lateral length features of the mouth relative to the nose; the ear and nose recognition region refers to the facial area between the ears and nose, which includes the left ear and nose recognition sub-region and the right ear and nose recognition sub-region, which mainly judges the facial area between the ears and nose and the distance features between the ears and nose; the facial recognition region mainly recognizes facial contour features, including the left facial contour recognition sub-region, the right facial contour recognition sub-region, the upper facial contour recognition sub-region, and the lower facial contour recognition sub-region.

[0046] The beneficial effects of this invention are as follows:

[0047] (1) This invention improves the accuracy and consistency of face recognition: By acquiring face images under different lighting and angles using a high-resolution camera, and combining preprocessing methods such as filtering and denoising, and grayscale normalization, the diversity and high quality of image data are ensured, providing a reliable foundation for subsequent processing. The ICP registration method is used to register and fuse feature points in multiple images of the same person acquired at different times and angles, so that key parts are aligned in different images, which significantly improves the accuracy and consistency of subsequent processing steps.

[0048] (2) Training and evaluation of the optimization model of this invention:

[0049] Generating high-quality 3D facial images: By converting 2D face images into 3D points in the camera coordinate system and performing rotation operations using a rotation matrix, rotated face profile data from multiple angles was generated, providing a rich data foundation for 3D reconstruction. A 3D reconstruction model was established using PointNet++ and its extended structures combined with an autoencoder framework. This model can extract compact feature representations from 3D point clouds and utilize these features to reconstruct 3D facial images with detail and geometric structure, improving the accuracy and realism of the reconstruction.

[0050] Adapting to the complex environment of underground coal mines: Addressing the issue that underground coal mine workers may wear masks, safety helmets, or have their faces obscured by dirt, the 3D facial image is divided into several recognition regions, and feature recognition and matching are performed on each region separately. This effectively improves the accuracy and robustness of recognition in complex environments. Identifying the corresponding identity information through individual recognition regions eliminates the need to stitch together images from all regions, simplifying the recognition process and improving efficiency. This solution can accurately and quickly identify underground coal mine workers, contributing to improved safety management and overall work efficiency in coal mining enterprises.

[0051] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0052] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0053] Figure 1 The flowchart of the three-dimensional face recognition method for underground coal mines provided by the present invention is shown. Detailed Implementation

[0054] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0055] Please see Figure 1 This invention provides a three-dimensional face recognition method for coal mines, specifically including the following steps:

[0056] S1: Acquire first face images using a high-resolution camera under different lighting and angles to ensure image diversity and coverage; the image preprocessing methods include filtering and noise reduction, grayscale normalization, edge detection, or size normalization; through image processing, output a relatively clean data image, eliminating a large number of irrelevant areas; the face matching model method is the ICP registration method, which involves geometrically transforming the same face images obtained at different times, different image blocks, and different shooting angles to make the corresponding image points completely superimposed in position and orientation.

[0057] Detailed calculation steps:

[0058] 1) Initialization: Among N facial images of the same face taken at different times and angles, a reference image I0 is selected as the baseline, and feature point set P0 is extracted from this image; for the other N-1 images I1, I2, ..., I... N-1 Extract the corresponding feature point sets P1, P2, ..., P respectively. N-1 .

[0059] 2) Registration process: For each pair of feature point sets P0 and P... i Perform the following steps:

[0060] 2.1) Find the nearest neighbor: For each point p0 in P0, find the nearest neighbor. i The nearest point p in the middle i It can be measured using Euclidean distance:

[0061] p i =arg min||p0-q||,q∈P i

[0062] 2.2) Calculate the transformation matrix: Based on the found nearest neighbor pair (p0, p... i Calculate the optimal rigid body transformation X i Rigid body transformation X i It typically includes a rotation matrix R. i Translation vector t i This can be solved using the least squares method:

[0063]

[0064] Where R is the rotation matrix and t is the translation vector. The rotation matrix is ​​calculated by first centering each point set, then calculating its covariance matrix, performing singular value decomposition (SVD) on the covariance matrix, and obtaining the rotation matrix from the SVD decomposition.

[0065] 2.3) Applying the transformation: Applying the calculated transformation R i and t i Applied to feature point set Pi This yields a new set of points:

[0066] P i ′=R i ·P i +t i

[0067] 3) Iterative update: Repeat step 2) until the feature point set P is reached. i The alignment error with P0 is less than a certain threshold or the maximum number of iterations is reached.

[0068] By following the steps above, feature points in multiple images of the same person collected at different times and angles can be registered and fused, aligning key features such as eyes, nose, and mouth in different images to obtain a second face image. This improves the accuracy and consistency of subsequent processing steps.

[0069] S2: The preprocessed second face image is divided into training, validation, and test sets according to a certain ratio. The training set is used for model learning, the validation set is used for model tuning, and the test set is used to evaluate model performance. Furthermore, feature point cloud data for constructing a 3D facial image is generated based on the processed second face image. Specifically, the method is as follows:

[0070] Given that the pixel value of each pixel in the second face image is I(x,y) and the initial depth value of each pixel is z(x,y), a virtual camera is set up based on the second face image, with an intrinsic parameter matrix of K and a focal length f relative to the second face image. x ,f y The center point of the image under the camera is c. x ,c y Then, each pixel (x, y) in the second face image is converted into a three-dimensional point (X, Y) in the camera coordinate system using the following formula. c ,Y c Z c ):

[0071]

[0072] Z c =z(x,y)

[0073] Then, rotate each point using a rotation matrix R. Assume the rotation matrix R contains three rotation angles θ. x ,θ y ,θ z Then the rotation matrix can be expressed as:

[0074] R = R z (θ z )·R y (θy )·R x (θ x )

[0075] Among them, R x ,R y ,R z These are rotation matrices around the x, y, and z axes, respectively, and can be represented as:

[0076]

[0077] The rotation matrix R can be used to rotate each point in the camera coordinate system:

[0078]

[0079] Among them, (X) w ,Y w Z w () is a point in the world coordinate system.

[0080] Specifically, step S2 generates rotated face profile data from four angles, and selects 30 different face data with different poses (10°, 30°, and 45°) as a test set, and calculates the comparison error between the test set and the reference pose face data.

[0081] S3: Establish a 3D reconstruction model from the 3D point cloud (X w ,Y w Z w Features are extracted from the data and a 3D facial image is constructed. The 3D reconstruction model employs PointNet++ and its extended architecture, combined with an autoencoder framework. PointNet++ is a deep learning model for processing point cloud data. It extends the original PointNet model by introducing local feature aggregation to enhance the understanding of the local structure of point clouds. When combined with an autoencoder, this model can extract compact feature representations from 3D point clouds and utilize these features to reconstruct 3D mesh or voxel models.

[0082] The encoder, primarily composed of PointNet++, is responsible for extracting features from the input 3D point cloud. This process is divided into multiple layers, each capturing geometric information from the point cloud through local feature aggregation. Specifically, each point interacts with its neighbors to compute its local feature representation. These local features are aggregated across multiple layers to ultimately obtain a global feature representation of the entire point cloud.

[0083] The decoder is responsible for converting the feature representations extracted by the encoder into a 3D mesh model or voxel model. This process is typically achieved through a series of deconvolution or upsampling operations, which gradually restore the feature representations to points or voxels in 3D space. During decoding, the model uses the features extracted by the encoder to generate a 3D facial image with detail and geometric structure.

[0084] Combining the encoder and decoder forms an autoencoder framework. In this framework, the model first extracts features from the input 3D point cloud using the encoder, and then uses the decoder to reconstruct these features into a 3D mesh model or voxel model. Through training, the model can learn how to effectively extract useful information from the point cloud and use this information to generate realistic 3D facial images.

[0085] The training process involves a large amount of 3D point cloud data and corresponding 3D mesh or voxel models. This data is typically acquired using a 3D scanner or other acquisition devices. During training, the model attempts to minimize reconstruction error, aiming to generate a 3D model that is as close as possible to the real 3D model. Through continuous iteration and optimization, the model gradually learns how to accurately reconstruct 3D facial images from point clouds.

[0086] S4: The trained 3D reconstruction model is used to identify the target faces of underground coal mine workers. Because underground coal mine workers typically wear masks, safety helmets, and noise-canceling headphones, or their faces may be obscured by dirt or coal dust due to environmental factors, recognition may fail, leading to misjudgments and impacting overall work efficiency.

[0087] Based on this, after obtaining the reconstructed three-dimensional facial image according to the aforementioned steps, the three-dimensional facial image is divided into several recognition regions, which include at least the eye recognition region, nose recognition region, mouth recognition region, eye and nose recognition region, mouth and nose recognition region, ear and nose recognition region, and face recognition region. The eye recognition region includes a left eye recognition sub-region and a right eye recognition sub-region, which respectively identify and match features of the left and right eyes; the nose recognition region identifies and matches nasal features; the mouth recognition region identifies and matches mouth features; the eye-nose recognition region refers to the facial area between the eyes and the nose, and includes a left eye nose recognition region and a right eye nose recognition region, which mainly judge the obvious features of the facial area between the eyes and the nose, as well as the distance features between the nose and the eyes; the mouth-nose recognition region refers to the facial area between the mouth and the nose, which mainly judges the obvious features of the facial area between the nose and the mouth, as well as the distance features between the nose and the mouth, and the lateral length features of the mouth relative to the nose; the ear-nose recognition region refers to the facial area between the ears and the nose, including a left ear-nose recognition sub-region and a right ear-nose recognition sub-region, which mainly judges the facial area between the ears and the nose, as well as the distance features between the ears and the nose; the face recognition region mainly recognizes facial contour features, including a left face contour recognition sub-region, a right face contour recognition sub-region, an upper face contour recognition sub-region, and a lower face contour recognition sub-region.

[0088] In the established database, the facial information of each person is pre-stored and split in the manner described above. Each split part carries corresponding identity information. That is, it is no longer necessary to stitch together the images of all regions. The identity information corresponding to a region can be known simply by recognizing the individual region.

[0089] Therefore, when performing facial recognition on workers in the mine, it is only necessary to select the clear facial features of the personnel for image reconstruction, and then match the features of a portion of the reconstructed clear facial image with the corresponding recognition area. Once the match is successful, the facial recognition result is given directly based on the identity information in the corresponding recognition area.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A three-dimensional face recognition method for use in underground coal mines, characterized in that, The method specifically includes the following steps: S1: Use a high-resolution camera to capture diverse and wide-coverage original face images under different lighting and angles, and preprocess them to output clean data images; then use the ICP registration method to register and fuse the same face images captured under different conditions to obtain a second face image. S2: Divide the second face image into training set, validation set and test set according to the proportion, and use them for model learning, optimization and performance evaluation respectively; at the same time, generate feature point cloud data for building three-dimensional facial images from the second face image, including rotating face profile data from multiple angles. S3: Establish a 3D reconstruction model, using PointNet++ and its extended structure combined with an autoencoder framework to extract features from 3D point clouds and reconstruct 3D facial images; through training, the model learns to extract useful information from point clouds and generate realistic 3D facial images. S4: The trained 3D reconstruction model is used to identify the target faces of underground coal mine workers. The 3D facial image is divided into several recognition regions, and feature recognition and matching are performed separately. In the information database, the facial information of each person is stored separately according to the recognition region, with identity information. During recognition, only clear regions need to be selected for image reconstruction and feature matching, and the face recognition result can be given directly.

2. The three-dimensional face recognition method for underground coal mines according to claim 1, characterized in that, In step S1, the original face image is preprocessed, specifically including: filtering and denoising, grayscale normalization, edge detection or size normalization.

3. The three-dimensional face recognition method for underground coal mines according to claim 1, characterized in that, In step S1, the ICP registration method involves geometrically transforming the same face images obtained at different times, in different image blocks, or from different shooting angles to ensure that corresponding image points are completely superimposed in position and orientation; the specific calculation steps are as follows: 1) Initialization: Among N facial images of the same face taken at different times or from different angles, a reference image I0 is selected as the baseline, and the feature point set P0 is extracted from this image; for the other N-1 images I1, I2, ..., I... N-1 Extract the corresponding feature point sets P1, P2, ..., P respectively. N-1 ; 2) Registration process: For each pair of feature point sets P0 and P... i Perform the following steps: 2.1) Find the nearest neighbor: For each point p0 in P0, find the nearest neighbor. i The nearest point p in the middle i Use Euclidean distance as a metric: p i =argmin||p0-q||,q∈P i 2.2) Calculate the transformation matrix: Based on the found nearest neighbor pair (p0, p... i Calculate the optimal rigid body transformation X i Rigid body transformation X i Includes rotation matrix R i Translation vector t i Solve using the least squares method: Where t is the translation vector; R is the rotation matrix. The rotation matrix is ​​calculated by first centering each point set, then calculating its covariance matrix, performing singular value decomposition (SVD) on the covariance matrix, and obtaining the rotation matrix from the SVD decomposition. 2.3) Applying the transformation: Applying the calculated transformation R i and t i Applied to feature point set P i This yields a new set of points: P i ′=R i ·P i +t i 3) Iterative update: Repeat step 2) until the feature point set P is reached. i The alignment error with P0 is less than a certain threshold or the maximum number of iterations is reached.

4. The three-dimensional face recognition method for underground coal mines according to claim 3, characterized in that, In step S2, feature point cloud data for building a 3D facial image is generated from the second face image. Specifically, this includes: given that the pixel value of each pixel in the second face image is I(x,y) and the initial depth value of each pixel is z(x,y), a virtual camera is set up based on the second face image, with its intrinsic parameter matrix being K and its focal length relative to the second face image being f. x ,f y The center point of the image under the camera is c. x ,c y Then, each pixel (x, y) in the second face image is converted into a three-dimensional point (X, Y) in the camera coordinate system using the following formula. c ,Y c Z c ): Z c =z(x,y) Then, rotate each point using a rotation matrix R; assuming the rotation matrix R contains three rotation angles θ. x ,θ y ,θ z Then the rotation matrix is ​​expressed as: R=R z (i z )·R y (i y )·R x (i x ) Among them, R x ,R y ,R z These are rotation matrices around the x, y, and z axes, respectively, and are expressed as: Rotate each point in the camera coordinate system using the rotation matrix R: Among them, (X) w ,Y w Z w () is a point in the world coordinate system.

5. The three-dimensional face recognition method for underground coal mines according to claim 1, characterized in that, In step S3, the established 3D reconstruction model specifically includes: The encoder part is composed of PointNet++, which is responsible for extracting features from the input 3D point cloud. This feature extraction process is divided into multiple layers, each layer capturing the geometric structure information in the point cloud through local feature aggregation. Specifically, each point interacts with its neighboring points to calculate the local feature representation of that point. These local features are aggregated at multiple layers to finally obtain the global feature representation of the entire point cloud. The decoder is responsible for converting the feature representations extracted by the encoder into a 3D mesh model or voxel model. This conversion process is achieved through a series of deconvolution operations or upsampling operations, which gradually restore the feature representations to points or voxels in 3D space. During the decoding process, the model uses the features extracted by the encoder to generate a 3D facial image with details and geometric structure. Combining the encoder and decoder forms an autoencoder framework. In this framework, the model first extracts features from the input 3D point cloud through the encoder, and then restores these features into a 3D mesh model or voxel model through the decoder.

6. The three-dimensional face recognition method for underground coal mines according to claim 1, characterized in that, In step S3, training the 3D reconstruction model specifically includes: during the training process, the model will try to minimize the reconstruction error, so that the generated 3D model is close to the real 3D model; through continuous iteration and optimization, the model gradually learns to accurately reconstruct 3D facial images from point clouds.

7. The three-dimensional face recognition method for underground coal mines according to claim 1, characterized in that, In step S4, the recognition area includes an eye recognition area, a nose recognition area, a mouth recognition area, an eye-nose recognition area, a mouth-nose recognition area, an ear-nose recognition area, and a face recognition area. The eye recognition area includes a left eye recognition sub-region and a right eye recognition sub-region, which respectively recognize and match features of the left and right eyes. The nose recognition area recognizes and matches nose features. The mouth recognition area recognizes and matches mouth features. The eye-nose recognition area refers to the facial area between the eyes and the nose, and consists of a left eye nose recognition area and a right eye nose recognition area, which determine the obvious features of the facial area between the eyes and the nose, as well as the nose. The distance features between the eyes; the mouth and nose recognition area refers to the facial area between the mouth and nose, judging the obvious features of the facial area between the nose and mouth, as well as the distance features between the nose and mouth, and the lateral length features of the mouth relative to the nose; the ear and nose recognition area refers to the facial area between the ears and nose, including the left ear and nose recognition sub-region and the right ear and nose recognition sub-region, judging the facial area between the ears and nose and the distance features between the ears and nose; the face recognition area recognizes facial contour features, including the left face contour recognition sub-region, the right face contour recognition sub-region, the upper face contour recognition sub-region, and the lower face contour recognition sub-region.

Images