A remote intelligent attendance method
Through the remote intelligent attendance method, the deep camera and face and body recognition models are used to solve the problems of inaccurate identity recognition and cheating in traditional attendance methods, and a more efficient and accurate employee check-in process is achieved.
Patent Information
- Application Number
- CN202410987312.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-23
- Publication Date
- 2024-12-10
- Estimated Expiration
- 2044-07-23
AI Technical Summary
Traditional attendance methods are prone to problems such as others taking the clock in instead of others and not checking in other places before they arrive, resulting in inaccurate identity identification and frequent cheating.
Remote intelligent attendance method is adopted to obtain human point cloud data and facial image data through high-precision depth cameras, and identity recognition is used to use facial recognition models and human model triangular networks to ensure the accuracy of the check-in process.
It effectively avoids cheating during attendance, improves the efficiency and accuracy of identity identification, and ensures the authenticity and reliability of employee attendance records.
Smart Images
Figure CN118521284B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of identity recognition, and in particular relates to a remote intelligent attendance method. Background Art
[0002] With the continuous advancement of technology, using attendance systems to manage the attendance of formal employees has become a common means for modern enterprises. The objects of this type of attendance system are mainly formal employees who are subject to standardized management by the enterprise. When checking attendance, each employee must strictly follow the company's prescribed start and end time. The company will determine each employee's attendance within the preset attendance cycle (including lateness, leave and overtime, etc.) based on each employee's attendance time, and use the attendance as the basis for employee salary settlement.
[0003] However, the traditional attendance method is mostly punch-in attendance or manual attendance. However, many companies' attendance mechanisms are prone to problems such as others punching in on behalf of others, and people punching in at other places when they are not there.
[0004] Based on this, there is an urgent need for an intelligent attendance method to solve the above problems. Summary of the invention
[0005] The purpose of the present invention is to provide a remote intelligent attendance method, which performs face and body recognition respectively through a face recognition model and a human body model triangular network, and determines the user identity based on the recognized face and body for sign-in attendance, thereby solving the problem that the existing enterprise clocking-in method is insufficient and easy to cheat.
[0006] To solve the above technical problems, the present invention is achieved through the following technical solutions:
[0007] The present invention is a remote intelligent attendance method, comprising the following steps:
[0008] Step S1: Scan the human body using a high-precision depth camera to obtain point cloud data of the human body surface and facial image data;
[0009] Step S2: constructing a human point cloud database and a face database;
[0010] Step S3: the attendance server trains the face recognition model and the human body model triangulation network;
[0011] Step S4: multiple high-precision depth cameras are installed at the enterprise's clock-in point to perform body scans on people entering the enterprise;
[0012] Step S5: Obtain the facial image of the person entering the enterprise and input it into the trained face recognition model to determine the identity of the person;
[0013] Step S6: The human body model triangulation network registers the point cloud data on the human body surface and calculates the best parameters for fitting the point cloud;
[0014] Step S7: setting parameter k to solve the body shape information of the corresponding position, obtaining the body shape information of the corresponding position and comparing it with the body shape of the person obtained by face recognition;
[0015] Step S8: If the comparison is successful, the normal attendance is completed; if the comparison fails, the personnel are notified that the attendance is not completed;
[0016] In step S6, the optimal parameters for fitting the point cloud are calculated, and the specific steps are:
[0017] Step S61: Under the condition that the positions of the fixed point and the moving point are determined, the least squares sum of the Laplace change values of all points is minimized, that is:
[0018] ;
[0019] In the formula, are the Laplace coordinates of the fixed points, is the Laplace coordinate of the moving point;
[0020] Step S62: Laplace transform equation Convert to ; Where K is a linear change matrix composed of human body parameters;
[0021] Step S63: Solving the formula ;
[0022] Step S64: According to the root-finding formula , calculate the final k value;
[0023] Step S65: scaling to make the size of the grid in step S3 consistent with that of the point cloud, and performing iterative registration.
[0024] As a preferred technical solution, in step S3, the preset human body model triangulation network needs to use Laplace deformation to set deformation parameters for various parts of the human body model, such as setting waist circumference deformation parameters, leg circumference deformation parameters, arm circumference deformation parameters, etc. to control mesh deformation. Since the person being scanned is required to keep his body fixed during scanning, the change of his body details hardly involves rotation operation, so we use Laplace deformation to realize parameterized deformation of the model;
[0025] In Laplace deformation, there are three types of points: fixed points, moving points, and points to be solved. These moving points are controlled by these parameters. In turn, these parameters represent a certain data of the human body. The points at the waist are the moving points controlled by the waist parameters. Due to the simplicity of matrix calculation, we regard the increase in waist circumference as a linear transformation. The coordinates of the points at the waist will be linearly related to the waist parameters. In the program, the relationship between the points is set to center expansion, that is, the points at the waist will be centered on the average value of their coordinates and scaled proportionally to the waist parameters. The higher the parameter, the larger the waist. The coordinates of other non-fixed points are Laplace deformed to obtain the final human body model.
[0026] As a preferred technical solution, in step S62, is an overdetermined linear equation system to be solved, where x is a vector composed of the coordinates of the points to be solved, each row of A represents a component of the Laplace coordinates of a component of a point, and b is the original Laplace coordinates and the coordinates of the known fixed points and their weights.
[0027] As a preferred technical solution, in step S64, according to the root-finding formula , the transformation solution can be obtained Then, we can derive the equation:
[0028] ;
[0029] Wherein, k is the corresponding parameter of the optimal body position, and the final value of k can be obtained by solving it; the parameter k is proportional to the body shape of the human body, and the body shape information of the corresponding position can be obtained by dividing the parameter k by the body shape proportion.
[0030] As a preferred technical solution, in step S65, the specific steps of iteratively making the size of the grid consistent with that of the point cloud are as follows:
[0031] Step R1: For each point in the grid, find the nearest point in the point cloud, then regard these points as x0 and solve the current optimal corresponding parameter k;
[0032] Step R2: Perform Laplace deformation on the parameter k introduced into the grid to obtain a new grid to replace the initial grid;
[0033] Step R3: Execute steps R1 and R2 repeatedly until the registration effect is achieved.
[0034] As a preferred technical solution, in step S5, the recognition process of the face recognition model is as follows:
[0035] Step S51, image preprocessing: modify the size of the face image taken by the high-precision depth camera, and adjust the image size to a suitable size for face recognition; due to the influence of factors such as lighting and imaging equipment, the image has random noise and blur interference, and the collected image needs to be denoised. The application document uses median filtering to denoise, which can effectively remove digital video noise, especially color image noise. At the same time, the image is grayed to remove color information, which is conducive to face recognition.
[0036] Step S52, face detection: face detection is performed through Haar features, and discriminative features are generated for face detection. For edges, vertices and thin lines in the image, discriminative features can be generated to achieve face detection;
[0037] Step S53, training data: using the train function of face recognition and the label array, identify the ID name of the face, and determine the identity of the person to be recognized based on the ID name. In this process, each person is divided into the same number of data sets for training, and two lists of face samples and ID names are established. Traverse the images in the list, convert the images into arrays and obtain the ID of each image, and finally obtain the loop object;
[0038] Step S54, LBPH face recognition: A confidence score is used to measure the difference between the recognized face and the real face, expressed as a numerical value, where 0 represents a perfect match, indicating that there is no difference between the recognized face and the original model. The higher the confidence score, the greater the difference between the two.
[0039] As a preferred technical solution, in step S51, the image preprocessing process is as follows:
[0040] Step S511, Gaussian filtering noise reduction: Assume that a pixel point position is , whose gray value is , multiply each pixel and its neighborhood by the Gaussian matrix, and take the weighted average value as the final grayscale value;
[0041] Step S512, calculate the gradient value and gradient direction: the smoothed image is filtered using the Sobel kernel in the horizontal and vertical directions to obtain the horizontal gradient value And the vertical gradient value ;
[0042] Step S513, filtering non-maximum values: in the Gaussian filtering process, use a unified rule to filter non-edge points so that the edge width is as close to 1 as possible;
[0043] Step S514, upper and lower threshold edge detection: two thresholds are set, representing the maximum threshold and the minimum threshold respectively. A value greater than the maximum threshold is detected as an edge, and a value less than the minimum threshold is detected as a non-edge.
[0044] As a preferred technical solution, in step S52, during face detection, the face sample set is placed in a comparison and screening module, and repeated face samples are removed by combining LBP feature coarse screening and depth feature fine screening; the face samples after removing the repeated faces are placed in a face recognition model for recognition, the recognition results are recorded in an Excel attendance record sheet, and the attendance record sheet is updated after each attendance is completed.
[0045] As a preferred technical solution, the LBP feature coarse screening uses deep convolution features to screen duplicate samples, extracts LBP features and uses Euclidean distance for feature comparison. When the feature distance of the two-by-two compared samples is less than the set threshold, the two samples are judged to be the same face. The specific calculation formula of the Euclidean distance is as follows:
[0046] ;
[0047] In the formula, Q and T represent the face samples screened for comparison, and the features of each face sample have n dimensions. and are the i-th dimension features of the original face sample and the i-th dimension features of the face sample to be screened respectively;
[0048] Then the LBP characteristic coarse screening formula is:
[0049] ; In the formula, is the center pixel of the image, is the gray value of the pth pixel, is the gray value of the center pixel.
[0050] As a preferred technical solution, the deep features are finely screened using a convolutional neural network; the convolutional neural network consists of four convolutional layers and one pooling layer, and is activated using a RELU activation function. The pooling layer uses maximum pooling with a step size of 2. The specific deep feature extraction method is as follows:
[0051] ;
[0052] In the formula, is the number of the current convolutional layer, Indicates The convolution kernel parameters in the convolution layer, including the convolution kernel format and convolution kernel size, is the bias of the current convolutional layer.
[0053] The present invention has the following beneficial effects:
[0054] The present invention recognizes faces and bodies respectively through a face recognition model and a human body model triangulation network, determines the identity of a user based on matching the recognized faces and bodies, and performs sign-in and attendance checking, thereby avoiding cheating in attendance checking and improving the efficiency and accuracy of identity recognition.
[0055] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0057] Figure 1 A flow chart of a remote intelligent attendance method of the present invention;
[0058] Figure 2 Flowchart for iteratively sizing the mesh to match the size of the point cloud. DETAILED DESCRIPTION
[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0060] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0061] In order to make the purpose, technical solutions and advantages of this application more clear, the following Figure 1-2 It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0062] See also Figure 1 As shown, the present invention is a remote intelligent attendance method, comprising the following steps:
[0063] Step S1: Scan the human body using a high-precision depth camera to obtain point cloud data of the human body surface and facial image data;
[0064] Step S2: constructing a human point cloud database and a face database;
[0065] Step S3: the attendance server trains the face recognition model and the human body model triangulation network;
[0066] Step S4: multiple high-precision depth cameras are installed at the enterprise's clock-in point to perform body scans on people entering the enterprise;
[0067] Step S5: Obtain the facial image of the person entering the enterprise and input it into the trained face recognition model to determine the identity of the person;
[0068] Step S6: The human body model triangulation network registers the point cloud data on the human body surface and calculates the best parameters for fitting the point cloud;
[0069] Step S7: setting parameter k to solve the body shape information of the corresponding position, obtaining the body shape information of the corresponding position and comparing it with the body shape of the person obtained by face recognition;
[0070] Step S8: If the comparison is successful, the normal attendance is completed; if the comparison fails, the personnel are notified that the attendance is not completed;
[0071] In step S6, the optimal parameters for fitting the point cloud are calculated, and the specific steps are as follows:
[0072] Step S61: Under the condition that the positions of the fixed point and the moving point are determined, the least squares sum of the Laplace change values of all points is minimized, that is:
[0073] ;
[0074] In the formula, are the Laplace coordinates of the fixed points, is the Laplace coordinate of the moving point;
[0075] Step S62: Laplace transform equation Convert to (1); where K is a linear variation matrix composed of human body parameters;
[0076] Step S63: Solving the formula ;
[0077] Step S64: According to the root-finding formula , calculate the final k value;
[0078] Step S65: scaling to make the size of the grid in step S3 consistent with that of the point cloud, and performing iterative registration.
[0079] In step S3, the preset human body model triangulation network needs to use Laplace deformation to set deformation parameters for each part of the human body model, such as setting waist circumference deformation parameters, leg circumference deformation parameters, arm circumference deformation parameters, etc. to control mesh deformation. Since the person being scanned is required to keep his body fixed during scanning, the change of his body details almost does not involve rotation operation, so we use Laplace deformation to achieve parameterized deformation of the model;
[0080] In Laplace deformation, there are three types of points: fixed points, moving points, and points to be solved. These moving points are controlled by these parameters. In turn, these parameters represent a certain data of the human body. The points at the waist are the moving points controlled by the waist parameters. Due to the simplicity of matrix calculation, we regard the increase in waist circumference as a linear transformation. The coordinates of the points at the waist will be linearly related to the waist parameters. In the program, the relationship between the points is set to center expansion, that is, the points at the waist will be centered on the average value of their coordinates and scaled proportionally to the waist parameters. The higher the parameter, the larger the waist. The coordinates of other non-fixed points are Laplace deformed to obtain the final human body model.
[0081] The weights used in Laplace coordinate calculations are arccos weights. Since the deformation parameters (waist circumference, chest circumference, arm circumference, etc.) are set for the model in advance, as long as the model is deformed under the set parameters and the parameters that best fit the point cloud are found, the size of the human body data corresponding to these parameters can be reversely calculated through the parameters.
[0082] In step S62, formula (1) can be transformed into an overdetermined linear equation system to be solved: (2)
[0083] In formula (2), x is the vector composed of the coordinates of the point to be solved, each row of A represents the component of the Laplace coordinate of a component of a point, and b is the original Laplace coordinate and the coordinates of the known fixed point and its weight;
[0084] Since formula (2) is the equation for solving the Laplace transform, and in the application, what is ultimately solved is the human body parameters we set; therefore, b in formula (2) is an undetermined quantity, because the coordinates of those moving points change with the change of the parameters; however, since when setting the parameters, our parameters are all linear with the coordinates, so b here must also satisfy a certain linear relationship with the parameters; that is, when multiple parameters form a linear change matrix K, b can be expressed as .
[0085] For example, when the parameter has only one waistline k, the coordinates of the "moving points" controlled by all waistlines satisfy , and among the two parts that make up b, the Laplace coordinates of v[i] will not change, so b can also be linearly represented by k.
[0086] Then, the formula in step S62 can be obtained: (3).
[0087] In step S62, is an overdetermined linear equation system to be solved, where x is a vector composed of the coordinates of the points to be solved, each row of A represents a component of the Laplace coordinates of a component of a point, and b is the original Laplace coordinates and the coordinates of the known fixed points and their weights.
[0088] In step S64, according to the root-finding formula , the transformation solution can be obtained Then, we can derive the equation:
[0089] (4);
[0090] In the formula, k is the corresponding parameter of the optimal body position. The final value of k can be obtained by solving it. The parameter k is proportional to the human body shape. The parameter k is divided by the body shape ratio to obtain the body shape information of the corresponding position.
[0091] See also Figure 2 As shown, in step S65, the specific steps of iterating to make the size of the grid consistent with that of the point cloud are as follows:
[0092] Step R1: For each point in the grid, find the nearest point in the point cloud, then regard these points as x0 and solve the current optimal corresponding parameter k;
[0093] Step R2: Perform Laplace deformation on the parameter k introduced into the grid to obtain a new grid to replace the initial grid;
[0094] Step R3: Execute steps R1 and R2 repeatedly until the registration effect is achieved.
[0095] In step S5, the recognition process of the face recognition model is as follows:
[0096] Step S51, image preprocessing: modify the size of the face image taken by the high-precision depth camera, and adjust the image size to a suitable size for face recognition; due to the influence of factors such as lighting and imaging equipment, the image has random noise and blur interference, and the collected image needs to be denoised. The application document uses median filtering to denoise, which can effectively remove digital video noise, especially color image noise. At the same time, the image is grayed to remove color information, which is conducive to face recognition.
[0097] Step S52, face detection: face detection is performed through Haar features, and discriminative features are generated for face detection. For edges, vertices and thin lines in the image, discriminative features can be generated to achieve face detection;
[0098] Step S53, training data: using the train function of face recognition and the label array, identify the ID name of the face, and determine the identity of the person to be recognized based on the ID name. In this process, each person is divided into the same number of data sets for training, and two lists of face samples and ID names are established. Traverse the images in the list, convert the images into arrays and obtain the ID of each image, and finally obtain the loop object;
[0099] Step S54, LBPH face recognition: A confidence score is used to measure the difference between the recognized face and the real face, expressed as a numerical value, where 0 represents a perfect match, indicating that there is no difference between the recognized face and the original model. The higher the confidence score, the greater the difference between the two.
[0100] In step S51, the image preprocessing process is as follows:
[0101] Step S511, Gaussian filtering noise reduction: Assume that a pixel point position is , whose gray value is , multiply each pixel and its neighborhood by the Gaussian matrix, and take the weighted average value as the final grayscale value;
[0102] Step S512, calculate the gradient value and gradient direction: the smoothed image is filtered using the Sobel kernel in the horizontal and vertical directions to obtain the horizontal gradient value And the vertical gradient value ;
[0103] Step S513, filtering non-maximum values: in the Gaussian filtering process, use a unified rule to filter non-edge points so that the edge width is as close to 1 as possible;
[0104] Step S514, upper and lower threshold edge detection: two thresholds are set, representing the maximum threshold and the minimum threshold respectively. A value greater than the maximum threshold is detected as an edge, and a value less than the minimum threshold is detected as a non-edge.
[0105] In step S52, during face detection, the face sample set is placed in a comparison and screening module, and duplicate face samples are removed by combining LBP feature coarse screening and depth feature fine screening; the face samples after removing duplicate faces are placed in a face recognition model for recognition, and the recognition results are recorded in an Excel attendance record sheet, and the attendance record sheet is updated after each attendance is completed.
[0106] LBP feature coarse screening uses deep convolution features to screen duplicate samples. It extracts LBP features and uses Euclidean distance for feature comparison. When the feature distance of the two-to-two samples is less than the set threshold, the two samples are judged to be the same face. The specific calculation formula of Euclidean distance is as follows:
[0107] ;
[0108] In the formula, Q and T represent the face samples screened for comparison, and the features of each face sample have n dimensions. and are the i-th dimension features of the original face sample and the i-th dimension features of the face sample to be screened respectively;
[0109] Subtract each pixel in the image from the central pixel. If the value is greater than the central pixel, it is 1, otherwise it is 0. This will form an 8-bit binary number, which is the LBP value of the central pixel. Using LBP feature coarse screening can screen out significant repeated samples through the shallow features of the face, about 60%.
[0110] Then the LBP characteristic coarse screening formula is:
[0111] ; In the formula, is the center pixel of the image, is the gray value of the pth pixel, is the gray value of the center pixel; when extracting LBP features, each pixel in the image is subtracted from the center pixel, and the value greater than the center pixel is 1, otherwise it is 0, thus forming an 8-bit binary number, which is the LBP value of the center pixel. At this time, the dimension of the LBP feature extracted from the image is 2 8 =256 dimensions, with a lower dimension, using LBP feature coarse screening can filter out significant repeated samples through shallow features of the face.
[0112] The convolutional neural network is used for fine screening of deep features. The convolutional neural network consists of four convolutional layers and one pooling layer, and is activated by the RELU activation function. The pooling layer uses maximum pooling with a step size of 2. The specific deep feature extraction method is as follows:
[0113] ;
[0114] In the formula, is the number of the current convolutional layer, Indicates The convolution kernel parameters in the convolution layer, including the convolution kernel format and convolution kernel size, is the bias of the current convolutional layer.
[0115] After extracting deep features from the roughly screened face samples, we use the similarity measurement method to achieve comparison and screening. If the similarity is greater than the threshold, it is determined to be the same face, and the sample is removed at this time. The threshold value is determined by experiments. Through a simple neural network, further screening can be performed simply and quickly from the remaining 40% of duplicate samples, which can greatly improve the time efficiency of face recognition. However, there will still be face samples that are difficult to screen and will enter the recognition stage. However, compared with the face recognition method without screening, the efficiency has been significantly improved.
[0116] It is worth noting that in the above system embodiment, the various units included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.
[0117] In addition, those skilled in the art can understand that all or part of the steps in the above-mentioned embodiments can be completed by instructing related hardware through a program, and the corresponding program can be stored in a computer-readable storage medium.
[0118] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.
Claims
1. A remote intelligent attendance method, characterized in that: The steps include: Step S1: Scan the human body using a high-precision depth camera to obtain point cloud data of the human body surface and facial image data; Step S2: constructing a human point cloud database and a face database; Step S3: the attendance server trains the face recognition model and the human body model triangulation network; Step S4: multiple high-precision depth cameras are installed at the enterprise's clock-in point to perform body scans on people entering the enterprise; Step S5: Obtain the facial image of the person entering the enterprise and input it into the trained face recognition model to determine the identity of the person; Step S6: the human body model triangulation network registers the point cloud data on the human body surface and calculates the best parameters for fitting the point cloud; Step S7: setting parameter k to solve the body shape information of the corresponding position, obtaining the body shape information of the corresponding position and comparing it with the body shape of the person identified by face recognition; Step S8: If the comparison is successful, the normal attendance is completed; if the comparison fails, the personnel are notified that the attendance is not completed; In step S6, the optimal parameters for fitting the point cloud are calculated, and the specific steps are as follows: Step S61: Under the condition that the positions of the fixed point and the moving point are determined, the least squares sum of the Laplace changes of all points is minimized, that is: ； In the formula, are the Laplace coordinates of the fixed points, is the Laplace coordinate of the moving point; Step S62: Laplace transform equation Convert to ; Where K is a linear change matrix composed of human body parameters; Step S63: Solving the formula ; Step S64: According to the root-finding formula , calculate the final k value; Step S65: scaling to make the size of the grid in step S3 consistent with that of the point cloud, and performing iterative registration; In step S62, is an overdetermined linear equation system to be solved, where x is a vector of coordinates of the points to be solved, each row of A represents a component of the Laplace coordinates of a component of a point, and b is the original Laplace coordinates and the coordinates of the known fixed points and their weights; In step S64, according to the root-finding formula , the transformation solution can be obtained Then, we can derive the equation: ； Wherein, k is the corresponding parameter of the optimal position of the body, and the final value of k can be obtained by solving it; the parameter k is proportional to the body shape of the human body, and the body shape information of the corresponding position can be obtained by dividing the parameter k by the body shape ratio; In step S65, the specific steps of iterating to make the size of the grid consistent with that of the point cloud are as follows: Step R1: For each point in the grid, find the nearest point in the point cloud, and then treat these points as , solve the current optimal corresponding parameter k; Step R2: Perform Laplace deformation on the parameter k introduced into the grid to obtain a new grid to replace the initial grid; Step R3: Execute steps R1 and R2 repeatedly until the registration effect is achieved.
2. A remote intelligent attendance method according to claim 1, characterized in that: In step S3, the preset human body model triangulated network needs to be deformed and parameterized for each part of the human body model through Laplace deformation.
3. A remote intelligent attendance method according to claim 1, characterized in that: In step S5, the recognition process of the face recognition model is as follows: Step S51, image preprocessing: modify the size of the face image taken by the high-precision depth camera to adjust the image size to a suitable size for face recognition; Step S52, face detection: performing face detection through Haar features, generating discriminative features for face detection; Step S53, training data: using the train function of face recognition and the label array, identify the ID name of the face, and determine the identity of the recognized person according to the ID name; Step S54, LBPH face recognition: a confidence score is used to measure the difference between the recognized face and the real face.
4. A remote intelligent attendance method according to claim 3, characterized in that: In step S51, the image preprocessing process is as follows: Step S511, Gaussian filtering noise reduction: Assume that a pixel point position is , whose gray value is , multiply each pixel and its neighborhood by the Gaussian matrix, and take the weighted average value as the final grayscale value; Step S512, calculate the gradient value and gradient direction: the smoothed image is filtered using the Sobel kernel in the horizontal and vertical directions to obtain the horizontal gradient value And the vertical gradient value ; Step S513, filtering non-maximum values: in the Gaussian filtering process, use a unified rule to filter non-edge points so that the edge width is as close to 1 as possible; Step S514, upper and lower threshold edge detection: two thresholds are set, representing the maximum threshold and the minimum threshold respectively. A value greater than the maximum threshold is detected as an edge, and a value less than the minimum threshold is detected as a non-edge.
5. A remote intelligent attendance method according to claim 4, characterized in that: In step S52, during face detection, the face sample set is placed in a comparison and screening module, and repeated face samples are removed by combining LBP feature coarse screening and depth feature fine screening; the face samples after removing repeated faces are placed in a face recognition model for recognition, and the recognition results are recorded in an Excel attendance record table, and the attendance record table is updated after each attendance is completed.
6. A remote intelligent attendance method according to claim 5, characterized in that: The LBP feature coarse screening uses deep convolution features to screen duplicate samples. The LBP feature is extracted and the Euclidean distance is used for feature comparison. When the feature distance of the two-to-two compared samples is less than the set threshold, the two samples are judged to be the same face. The specific calculation formula of the Euclidean distance is as follows: ； In the formula, Q and T represent the face samples screened for comparison, and the features of each face sample have n dimensions. and are the i-th dimension features of the original face sample and the i-th dimension features of the face sample to be screened respectively; Then the LBP characteristic coarse screening formula is: ; In the formula, is the center pixel of the image, is the gray value of the pth pixel, is the gray value of the center pixel.
7. A remote intelligent attendance method according to claim 5, characterized in that: The deep features are finely screened using a convolutional neural network; the convolutional neural network consists of four convolutional layers and one pooling layer, and is activated using a RELU activation function. The pooling layer uses maximum pooling with a step size of 2. The specific deep feature extraction method is as follows: ； In the formula, is the number of the current convolutional layer, Indicates The convolution kernel parameters in the convolution layer, including the convolution kernel format and convolution kernel size, is the bias of the current convolutional layer.