Distributed storage method for face image data

Abstract

The present invention discloses a distributed storage method for facial image data, which belongs to the field of artificial intelligence and aims to achieve accurate driver identity authentication and status monitoring by real-time capture and distributed processing of driver facial images. It includes: real-time capture of driver facial images by front-end acquisition equipment; using edge computing nodes to determine the lighting conditions of the captured facial images, performing dynamic normalization processing according to the lighting conditions, adaptively graying the facial images based on the lighting conditions, and uploading the processed facial images to a data center after scaling and cutting, then performing feature extraction and constructing a feature set; based on the feature set, performing preliminary processing, and then creating redundant backups and establishing indexes in the data center through distributed storage; the personalized processing unit compares the current driver feature set with the feature set stored in the data center to determine whether the driver's facial image matches to determine the control authority of the vehicle.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and more specifically to a distributed storage method for face image data. Background Art

[0002] In today's automotive industry, intelligent driving technology is developing rapidly, aiming to provide a safer and more convenient driving experience. Among them, face recognition technology, as an important biometric means, has been widely used in aspects such as driver identity authentication and personalized settings adjustment. However, with the continuous increase in the amount of data, how to efficiently and securely store and manage these facial data has become an urgent problem to be solved. Distributed storage technology has attracted attention due to its high reliability, high scalability and high performance. By distributing and storing data on multiple independent devices, distributed storage can not only improve the disaster recovery capability of data, but also dynamically expand storage capacity according to demand. This is of great significance for intelligent driving systems that need to process a large amount of driver facial data. In the intelligent driving scenario, the driver's facial image data can be collected in real time through the in-car camera and processed and analyzed using advanced face recognition algorithms. This data is not only used for identity authentication, but also for the configuration of vehicle control permissions, such as automatically adjusting the seat position, rearview mirror angle, etc., to enhance the driving experience. At the same time, by identifying the driver's emotional state, the system can also make corresponding environmental adjustments, such as playing music or activating the seat massage function, to further increase driving comfort and safety. In order to protect the driver's personal privacy and data security, it is particularly important to use distributed storage technology to encrypt, store and manage facial image data. This not only prevents data leakage and unauthorized access, but also ensures compliance with relevant laws and regulations. In addition, the high availability and scalability of the distributed storage system also ensure that the data processing speed can be maintained quickly even with large amounts of data, which is crucial for intelligent driving systems that require instant feedback.

[0003] The invention patent with announcement number CN114063909B discloses an intelligent distributed storage method and system for image data, including an image storage module, an image restoration module and an image reading and parsing module; combining the spatial redundancy characteristics of the image data itself, on the basis of the existing redundancy mechanism, combined with intelligent algorithms to provide data security protection.

[0004] Although the prior art improves the spatial availability of image data in distributed storage and improves the ability to repair data after image storage data blocks are lost, it still fails to capture and distribute driver facial images in real time and achieve accurate identity authentication and status monitoring. Therefore, in order to overcome these limitations, the present invention proposes a distributed storage method for facial image data. Summary of the invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a distributed storage method for facial image data, which solves the problem of real-time capture and distributed processing of driver facial images, and realizes accurate identity authentication and status monitoring. The driver's facial image is captured in real time by the front-end device, and the edge computing node is used to perform dynamic normalization processing such as graying, scaling and cropping before uploading to the data center. The data center extracts local and global features of the normalized image and constructs a feature set. A distributed storage architecture is adopted, and the replica factor is configured according to the importance of the feature for redundant backup, and data access requests are dispersed through load balancing technology. Combined with the network connection status, remote or local feature matching is selected for identity authentication to provide vehicle control authority. On the basis of identity authentication, the driver's status is continuously monitored to ensure driving safety.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A distributed storage method for face image data comprises the following steps:

[0008] Step S1: Capturing the driver's face image in real time through the front-end acquisition device;

[0009] Step S2: using the edge computing node to determine the lighting conditions of the captured face image, performing dynamic normalization processing according to the lighting conditions, adaptively graying the face image based on the lighting conditions, and uploading the processed face image to the data center after scaling and cropping;

[0010] Step S3: the data center performs feature extraction on the face image data after dynamic normalization processing, wherein the feature extraction includes local feature extraction and global feature extraction, and constructs a feature set;

[0011] Step S4: Based on the feature set, preliminary processing is performed, including feature set splitting, encoding, compression, encryption, and then redundant backup is created and indexing is established in the data center through distributed storage;

[0012] Step S5: The personalized processing unit compares the current driver feature set with the feature set stored in the data center to determine whether the driver's facial image matches to determine the control authority of the vehicle.

[0013] Specifically, the specific steps of dynamic normalization processing include:

[0014] Get a face image and determine the lighting conditions by calculating the statistics of the brightness distribution, contrast, and color saturation of the face image. The calculation formula is as follows:

[0015] ,in, is the comprehensive evaluation value of lighting conditions. is the mean brightness of the face image, is the standard deviation of face image brightness, is the color saturation of the face image, is the face image contrast, is a positive constant;

[0016] Configuring light thresholds and ,and Less than ,like , then the face image is judged to be in a low-light environment. If , then the face image is judged to be in a high-light environment, otherwise it is in a normal-light environment.

[0017] Specifically, the specific steps of local feature extraction include:

[0018] Identify key points in the image. For each detected key point, configure a window, divide the pixels in the window into sub-windows, and calculate the gradient histogram of each sub-window. Connect the histograms of all sub-windows to form a descriptor vector.

[0019] All descriptor vectors are stacked into a matrix, singular value decomposition is performed, and the eigenvectors corresponding to the largest k eigenvalues ​​are selected for data projection to achieve dimensionality reduction;

[0020] The distance between the descriptor vector to be matched and each descriptor vector in the face database is calculated, and the local facial feature label of the descriptor vector with the smallest distance is used as the label of the descriptor vector to be matched.

[0021] Specifically, the specific steps of global feature extraction include:

[0022] Perform global texture extraction on face images, apply neighborhood windows of different sizes to face images, perform multi-scale extraction, calculate the LBP eigenvalue of each window, merge the LBP eigenvalues ​​of all scales, and form a multi-scale LBP feature vector;

[0023] Perform global edge shape feature extraction on face images, divide the face images into multiple levels of regions, extract HOG features independently on each level of region, and merge features of all levels into a long vector.

[0024] Specifically, the steps of global texture extraction include:

[0025] Configure the neighborhood window size and apply the neighborhood window to each pixel with it as the center;

[0026] Calculate the LBP feature value of each window, that is: ,in, is the coordinate of the center pixel, It is The coordinates of the surrounding pixels, Representing coordinates The pixel value at Representing coordinates The pixel value at is the number of pixels in the neighborhood, is a step function, defined as:

[0027]

[0028] at this time, Pick When Greater than ,but The value is 1, if Less than or equal to ,but The value is 0;

[0029] Configuring the Offset Angle Threshold , for each pixel point, calculate its LBP value within the offset angle threshold range, and take its minimum value as the final LBP feature value of the pixel point, that is: ,in, is the rotation angle, and its value range is , is the rotation angle The LBP eigenvalue under ;

[0030] Merge the LBP eigenvalues ​​of all neighborhood window configurations and merge the LBP eigenvalues ​​obtained at each scale into a long vector.

[0031] Specifically, the specific steps of global edge shape feature extraction include:

[0032] Divide the face image into multiple hierarchical regions to form a multi-level grid structure;

[0033] For each level of the region, calculate the gradient amplitude and gradient direction of each pixel in the region. For any pixel in the region , the calculation formula of gradient amplitude and gradient direction is:

[0034]

[0035]

[0036] in, It's a pixel The gradient amplitude of It's a pixel The gradient direction of and It's a pixel Horizontal and vertical gradients;

[0037] The gradient direction is divided into interval directions. For each pixel point, its gradient amplitude is accumulated into the histogram of the corresponding interval according to its gradient direction, and the gradient direction histogram of each area is dynamically normalized.

[0038] Combine the gradient direction histograms in the regions of all levels into one long vector.

[0039] Specifically, the specific steps of feature set encryption include:

[0040] Based on the encoded and compressed data fragments, different encryption levels are used for the original face image, local features and global features, and multi-level encryption is performed separately, including:

[0041] The first level of encryption uses an encryption algorithm to perform a first-level encryption on the encoded and compressed original face image data fragment, generates a key and performs key encoding, which is used as a key feature vector and saved in the local feature vector;

[0042] The second level of encryption uses an encryption algorithm to perform secondary encryption on the local feature data fragments, generates a new key and encodes the key as a key feature vector, which is saved in the global feature vector;

[0043] The third level of encryption uses an encryption algorithm to perform three-level encryption on the global feature data fragment, generate the final key and encode the key as a key feature vector, and save it in a secure storage area.

[0044] Specifically, the specific steps of distributed storage include:

[0045] The data center performs deep processing on the received facial feature set, wherein the deep processing includes feature set splitting, feature set encoding, feature set compression, and feature set encryption;

[0046] According to the encryption level, respectively configure the replication factor for the original face image, the local features and the global features, and perform redundant backup according to the replication factor;

[0047] Configure a distributed storage architecture for the data center and perform distributed storage of the redundantly backed-up original face images, local features, and global features;

[0048] Through load balancing technology, data access requests are distributed to multiple data nodes in the distributed storage architecture to avoid overloading of a single node;

[0049] A local storage device is set up inside the vehicle to cache the normalized face images obtained by the edge computing node and the global features extracted by the data center;

[0050] Configuring Update Thresholds , Used to measure the time interval required for updating. Used to measure the memory capacity that needs to be updated. When the interval between local storage device cache updates is greater than When the local storage device has less memory capacity than When , the normalized face images cached locally are adaptively synchronized to the data center.

[0051] Specifically, the steps for identity verification include:

[0052] Detect whether the current vehicle network is connected. If the network is connected, send a request to the remote data center to obtain the real-time updated facial image feature set, perform complex feature matching on the captured driver's facial image features and the feature set provided by the data center, and determine whether the current driver already exists in the data center of the vehicle based on the matching results;

[0053] If the network is not connected, it will enter local verification mode, read the pre-cached normalized facial image and global features from the vehicle's local storage device, use local data for simple feature matching to determine the driver's identity, and determine whether the current driver already exists in the vehicle's local storage device based on the local matching results.

[0054] Specifically, simple feature matching includes:

[0055] Obtain the normalized face image of the current driver through the edge computing node, measure the similarity with the normalized face image in the local storage device, configure the similarity threshold, and output the face image identifier whose similarity measurement is greater than the similarity threshold;

[0056] When the number of output face image identifiers is 0: the access request of the current driver is rejected and the vehicle control authority is not provided;

[0057] When the number of output face image identifiers is 1: the recognition result is displayed through a visual interface, and the current driver is requested to confirm;

[0058] When the number of output facial image identifiers is greater than 1: for each output facial image identifier, obtain its corresponding global features, call the edge computing node to obtain the global features of the current driver's facial image, perform similarity measurement, output the facial image identifier with the highest similarity, and request the current driver to confirm.

[0059] Beneficial effects of the present invention:

[0060] 1. Dynamic normalization of captured facial images through edge computing nodes, including grayscale, scaling, and cropping, can effectively reduce the amount of data transmission and improve transmission efficiency. The data center extracts features of the facial image data after dynamic normalization, including local feature extraction and global feature extraction, and builds a feature set, which helps improve the accuracy and efficiency of face recognition. The splitting, encoding, compression, and encryption of feature sets, as well as the creation of redundant backups and indexing in the data center through distributed storage, can improve the reliability and security of data.

[0061] 2. The personalized processing unit compares the current driver feature set with the feature set stored in the data center to determine whether the driver's face image matches to determine the vehicle's control authority, thereby enhancing the user experience. During the identity authentication process, whether in online or offline mode, the driver can be effectively identified and the corresponding vehicle control authority can be provided, ensuring the continuity and personalization of the service.

[0062] 3. The application of distributed storage architecture and load balancing technology can avoid the overload problem of a single node and improve the response speed and stability of the system. The setting of local storage devices is used to cache the normalized face images obtained by edge computing nodes and the global features extracted by the data center, which can reduce the dependence on the central server and reduce network latency. The adaptive synchronization mechanism ensures that the normalized face images cached locally can be updated to the data center in a timely manner to maintain data consistency. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A flowchart of a distributed storage method for facial image data;

[0064] Figure 2 This is a flowchart of adaptive grayscale;

[0065] Figure 3 Flowchart for feature extraction;

[0066] Figure 4 Flowchart of the personalized processing unit. DETAILED DESCRIPTION

[0067] See also Figure 1 This embodiment introduces a distributed storage method for face image data, including:

[0068] Step S1: Capturing the driver's face image in real time through the front-end acquisition device;

[0069] In this embodiment, a high-resolution camera is installed inside the vehicle to ensure that the driver's facial details can be clearly captured. The real-time video streaming function of the camera is used to continuously monitor and capture the driver's facial image, providing a high-quality data source for subsequent feature extraction and identity verification.

[0070] Step S2: Use the edge computing node to dynamically normalize the captured face image and upload it to the data center, including grayscale, scaling, and cropping to eliminate the influence of factors such as lighting and posture;

[0071] In this embodiment, the edge computing node receives image data from the camera and performs preprocessing. First, the color image is converted into a grayscale image to reduce the computational complexity. Then, the image size is adjusted through scaling and cropping techniques to ensure that all images have a uniform size and ratio. This ensures that the subsequent feature extraction process is not interfered by external variables, improves the accuracy and reliability of recognition, and can run stably in various environments.

[0072] Preferably, the specific steps of dynamic normalization processing include:

[0073] See also Figure 2 , obtain the face image, and judge the lighting conditions by calculating the lighting statistics of the brightness distribution, contrast and color saturation of the face image. The calculation formula of the lighting statistics is as follows:

[0074]

[0075] in, is the comprehensive evaluation value of lighting conditions. is the mean brightness of the face image, and the calculation formula is:

[0076]

[0077] is the standard deviation of the brightness of the face image, and the calculation formula is:

[0078]

[0079] is the color saturation of the face image, and the calculation formula is:

[0080]

[0081]

[0082] is the contrast of the face image, and the calculation formula is:

[0083]

[0084] in, Is a face image The length and width of is the color channel of the face image , and the coordinates are Pixel value, color channel of face image The value range is R, G, B; Is a normal number used to prevent division by zero. The ratio of the brightness mean to the standard deviation reflects the average brightness of the face image relative to its range of variation. If the ratio is high, it means that the face image is brighter and uniform overall; if the ratio is low, it means that the face image is darker or there are large brightness fluctuations. The ratio of color saturation to contrast reflects the color saturation of the image relative to the contrast. If the ratio is high, it means that the image color is brighter and the contrast is obvious; if the ratio is low, it means that the image color is relatively simple or the contrast is insufficient. Nonlinear function It is used to adjust the influence of brightness standard deviation on comprehensive evaluation. When the brightness standard deviation is large, the function value is close to 1, emphasizing the influence of brightness change on lighting conditions; when the brightness standard deviation is small, the function value is close to 0, weakening the influence of brightness change. By adding the above two ratios and multiplying them by the nonlinear adjustment factor, a comprehensive evaluation value is obtained. , taking into account multiple factors such as brightness, color saturation and contrast, it can more comprehensively reflect the lighting conditions.

[0085] Configuring light thresholds and ,and Less than ,like , then the face image is judged to be in a low-light environment. If , then the face image is judged to be in a high-light environment, otherwise it is in a normal light environment. When in a low-light environment, the overall brightness of the face image is low, resulting in unclear facial features and increased recognition difficulty. When in a high-light environment, the face image will be overexposed, resulting in loss of facial details.

[0086] Adaptive grayscale is performed according to the lighting conditions, and the reference weight coefficient for grayscale of the face image under normal light environment is configured. The grayscale face image is obtained by weighted averaging through the reference weight coefficient, that is:

[0087]

[0088] in, is the grayscale face image, is the pixel value of the face image under the R channel, is the pixel value of the face image under the G channel, is the pixel value of the face image under channel B, , , They are the reference weight coefficients corresponding to the pixel values ​​under the R, G, and B channels. The weight coefficients are adaptively modified according to the lighting conditions to achieve the purpose of adaptive grayscale. The specific formula for adaptively modifying the weight coefficients is as follows:

[0089]

[0090]

[0091]

[0092] in, , , They are the reference weight coefficients corresponding to the pixel values ​​under the modified R, G, and B channels, respectively. is the comprehensive evaluation value of the average lighting conditions under normal lighting conditions. , , They are the corresponding modification parameters for the R, G, and B channels, with a value range of [0,1]. is a mapping function used to transform The value of is mapped to [0,1], and the grayscale coefficient of the face image is adaptively modified according to the lighting conditions of the face image, changing the influence of different channels on the grayscale image, so that the grayscale image can maintain the best visual effect and information integrity under various lighting conditions. Specifically, this process achieves dynamic adaptation to different lighting environments by adjusting the baseline weight coefficients of the R, G, and B channels, thereby optimizing the quality of the grayscale image.

[0093] By using the face detection algorithm, including the Haar feature classifier in OpenCV, the face position in the grayscale image is detected, the detected face boundary coordinates are calculated to determine the minimum rectangular area containing the entire face, and the original image is cropped according to the determined rectangular area, the face image in the rectangular area is retained, and the face image in the rectangular area is scaled according to the target size; in order to remove the background information of the driver's face image captured in real time by the front-end acquisition device, so that the subsequent feature extraction and recognition are more accurate. And according to the target size, the cropped face image is scaled to adapt to the required input size, and the processed image will have a uniform size and content, which is convenient for further analysis and comparison.

[0094] Step S3: The data center performs feature extraction on the face image data after dynamic normalization, wherein the feature extraction includes local feature extraction and global feature extraction, and constructs a feature set;

[0095] In this example, see Figure 3 , feature extraction is performed on the dynamically normalized image. The local feature extraction algorithm is used to identify the feature points of key areas such as the eyes, nose, and mouth. At the same time, the global feature extraction algorithm is used to analyze the texture and contour information of the entire face. These feature points are combined with the global features to construct a comprehensive face feature set, which provides a solid foundation for identity authentication.

[0096] Preferably, the specific steps of local feature extraction include:

[0097] Harris corner detector is used to identify key points in face images. For each detected key point, a window is configured, the pixels in the window are divided into sub-windows, and the gradient histogram of each sub-window is calculated. The histograms of all sub-windows are connected to form a descriptor vector.

[0098] Calculate the gradient magnitude and gradient direction of the face image, apply the Harris corner detection algorithm, and identify potential key points by calculating the interest value. The interest value is determined by the eigenvalue of the gradient covariance matrix. The interest value calculation formula is:

[0099]

[0100] in, is the value of interest, M is the gradient covariance matrix, is the determinant of the gradient covariance matrix, is the regulating factor, is the trace of the gradient covariance matrix.

[0101] Through non-maximum suppression, the key points with the maximum response are retained in the adjacent area of ​​each key point. The key points correspond to the corner points of facial features such as eyes, nose, and mouth. By identifying these key points, a basis is provided for subsequent feature extraction. In a certain range around each key point, that is, a 3x3 or 5x5 window, the key point with the maximum interest value is obtained as the final key point. With the key point as the center, a fixed-size window, such as 8x8, is selected, and the window is divided into multiple sub-windows. For example, an 8x8 window can be divided into 4 sub-windows. The gradient histograms of the pixels in each sub-window are calculated in 8 directions, and the gradient histograms of all sub-windows are connected to form a long vector as the descriptor vector of the key point. The descriptor vector contains the local feature information around the key point, including edge direction, texture, etc. In this way, the key point is associated with the local features around it, providing a basis for subsequent feature matching.

[0102] All descriptor vectors are stacked into a matrix, singular value decomposition is performed, and the eigenvectors corresponding to the largest k eigenvalues ​​are selected for data projection to achieve dimensionality reduction and obtain the local descriptor eigenvector;

[0103] Create an empty matrix. The number of columns in the matrix will be equal to the dimension of a single descriptor vector, and the number of rows will be equal to the number of key points. Iterate over the descriptor vector of each key point and add it to the matrix as a row of the matrix. After traversing all key points, stack the descriptor vectors of all key points into a matrix; perform singular value decomposition on the matrix and retain the top k largest eigenvalues ​​and their corresponding eigenvectors. Use these eigenvectors to project the original data, thereby achieving dimensionality reduction, reducing the dimension of the data, reducing the computational complexity, and retaining the main feature information of the data.

[0104] Calculate the distance between the local descriptor feature vector to be matched and each descriptor vector in the face database, and use the local face feature label of the descriptor vector with the smallest distance as the label of the local descriptor feature vector to be matched;

[0105] Calculate the Euclidean distance or other similarity measure between the local descriptor feature vector to be matched and each descriptor vector in the face database. Find the descriptor vector pair with the smallest distance and regard it as the best match. Taking the descriptor vector pair with the smallest distance as the best match means that the two descriptor vectors are closest in the feature space and have the highest similarity. By comparing the similarity between the local descriptor feature vector in the current image and the local descriptor feature vector stored in the face database, find the most similar pair of descriptor vectors, obtain the local facial feature label corresponding to the descriptor vector, and thus achieve label matching;

[0106] Preferably, the specific steps of global feature extraction include:

[0107] Perform global texture extraction on face images, apply neighborhood windows of different sizes to face images, perform multi-scale extraction, calculate the LBP eigenvalue of each window, merge the LBP eigenvalues ​​of all scales, and form a multi-scale LBP feature vector;

[0108] The global edge shape feature extraction is performed on the face image, the face image is divided into multiple levels of regions, the HOG feature values ​​are extracted independently in each level of the region, and the feature values ​​of all levels are merged into a multi-level HOG feature vector.

[0109] Preferably, the step of extracting global texture from the face image includes:

[0110] Configure the neighborhood window size and apply the neighborhood window to each pixel with it as the center;

[0111] Calculate the LBP feature value of each window, that is:

[0112] in, is the coordinate of the center pixel, It is The coordinates of the surrounding pixels, Representing coordinates The pixel value at Representing coordinates The pixel value at is the number of pixels in the neighborhood, excluding the center point, is a step function, defined as:

[0113]

[0114] at this time Pick When Greater than ,but The value is 1, if Less than or equal to ,but The value is 0;

[0115] Configuring the Offset Angle Threshold , for each pixel point, calculate its LBP value within the offset angle threshold range, and take its minimum value as the final LBP feature value of the pixel point, that is:

[0116]

[0117] in, is the rotation angle, and its value range is , exemplary, The value is 90°. is the rotation angle The LBP eigenvalue under ;

[0118] Merge the LBP eigenvalues ​​of all neighborhood window configurations, and merge the LBP eigenvalues ​​obtained at each scale into a long vector. Get an LBP feature vector that contains multiple scale information and is rotationally invariant. This feature vector can capture the texture information of the image at different scales, making the features more distinguishable. By minimizing the LBP value of each pixel at different rotation angles, the features are made robust to the rotation changes of the image. Combining the characteristics of multi-scale and rotational invariance, this feature vector can more comprehensively describe the texture information of the image, and is suitable for various application scenarios that require high robustness and accuracy.

[0119] Preferably, the specific steps of extracting global edge shape features include:

[0120] The face image is divided into multiple hierarchical regions, which gradually change from large to small, forming a multi-level grid structure, starting with a larger grid, such as 8x8 pixels, and then gradually reducing the size of the grid, such as 4x4 pixels, 2x2 pixels. This multi-level division method can capture information of different scales in the image. The multi-level regional division method makes the features more robust to subtle changes in the image and can maintain good performance in various scenarios;

[0121] For each level of the region, calculate the gradient amplitude and gradient direction of each pixel in the region. For any pixel in the region , the calculation formula of gradient amplitude and gradient direction is:

[0122]

[0123]

[0124] in, It's a pixel The gradient amplitude of It's a pixel The gradient direction of and It's a pixel Horizontal and vertical gradients;

[0125] The gradient direction is divided into interval directions, each interval represents a specific angle range. For each pixel point, its gradient amplitude is accumulated into the histogram of the corresponding interval according to its gradient direction, and the gradient direction histogram of each area is dynamically normalized to reduce the impact of illumination and contrast changes;

[0126] The gradient direction histograms in all levels of regions are merged into a long vector to form a HOG feature vector; through multi-level region division, precise gradient calculation, effective interval direction division, dynamic normalization processing and feature merging, the global edge shape feature extraction method can effectively capture the edge information and local shape information in the image.

[0127] Preferably, the specific steps of constructing the feature set include:

[0128] Construct a temporary database of feature sets, whose fields include: face image identifier, original face image, normalized face image, face image file path, face image acquisition time, local descriptor feature vector, local descriptor feature vector label, global multi-scale LBP feature vector, global multi-level HOG feature vector, which is used to store various information related to face features for subsequent processing and analysis.

[0129] The feature vectors, including the local descriptor feature vectors, the global multi-scale LBP feature vectors and the multi-level HOG feature vectors, as well as their feature information including the vector labels and file paths, are updated to the feature set database, and it is ensured that all the extracted feature data are systematically stored to facilitate subsequent identity authentication or recognition tasks.

[0130] After integrating all feature data into the database, data cleaning and optimization are performed to remove redundant data and standardize feature vectors to improve the efficiency and accuracy of data processing.

[0131] Establish indexes to optimize queries, establish indexes on key fields, including face image identifiers, local descriptor feature vector labels, etc., to speed up matching during identity authentication and optimize database query performance.

[0132] Step S4: Based on the feature set, preliminary processing is performed, including feature set splitting, encoding, compression, encryption, and then redundant backup is created and indexing is established in the data center through distributed storage;

[0133] In this embodiment, the extracted feature set temporary database is split and then encoded and compressed to reduce the space requirements for data transmission and storage. Then, the compressed data is encrypted to ensure data privacy and security. Through a distributed storage system, the encrypted feature set is stored in multiple data centers to create redundant backups to prevent data loss. At the same time, an index is established for each data fragment to improve the efficiency of data retrieval.

[0134] Preferably, the specific steps of encoding and compressing the feature set include:

[0135] Based on the feature set temporary database, the original face image, the normalized face image, the local descriptor feature vector, the global multi-scale LBP feature vector, and the global multi-level HOG feature vector are split into independent data segments for separate processing and storage.

[0136] Encode each data fragment and convert it into binary format. Use Python's pickle module to serialize and convert the data fragment into binary format.

[0137] Compress the encoded data fragments using the Gzip compression algorithm to ensure that the data takes up less space during transmission and storage. And use the distributed computing framework to process the encoding and compression tasks of multiple data fragments in parallel to further improve the processing speed;

[0138] Preferably, the specific steps of encrypting the feature set include:

[0139] Based on the encoded and compressed data fragments, the original face image, local descriptor feature vector, global multi-scale LBP feature vector, and global multi-level HOG feature vector are encrypted at multiple levels, including:

[0140] The first level of encryption uses a strong encryption algorithm, including AES-256 bit encryption, to encrypt the encoded and compressed original face image data fragments. A key is generated from the first level encryption process, and the key is encoded and converted into a hexadecimal string. The encoded key is used as a key feature vector and saved in the local descriptor feature vector. By encoding and compressing the original face image data and encrypting it in the first level, even if the data is intercepted during transmission, unauthorized third parties cannot directly interpret the image content, thereby protecting the user's privacy. The generated key feature vector is part of the local feature, and its security is enhanced because even if the local feature data is leaked, the original image cannot be decrypted without the correct key.

[0141] The second level of encryption uses an encryption algorithm to perform secondary encryption on the extracted local descriptor feature vector data fragments, including local area features such as eyes, nose, and mouth. A new key is generated from the secondary encryption process and encoded. The encoded new key is used as a key feature vector and saved in the global feature vector, including the global multi-scale LBP feature vector and the global multi-level HOG feature vector. The secondary encryption of local features provides an additional layer of protection for data security. This means that even if the local feature data fragments are leaked, the attacker still cannot obtain useful information because they have been encrypted. Each level of encryption generates and encodes new keys, which are independent of each other, increasing the difficulty of cracking because the attacker needs to obtain all keys at the same time to fully decrypt the data.

[0142] The third level of encryption uses the same encryption algorithm to perform three-level encryption on the extracted global feature data fragments. The final key is generated from the three-level encryption process and encoded. The encoded final key is saved in the secure storage area as a key feature vector to ensure its security and integrity. The three-level encryption of global features ensures that even if the first two levels of encryption are breached, the final key is still securely stored in the system's secure storage, providing the last line of defense for the entire system. Through this layered encryption strategy, the overall security of the system is significantly improved because each level of encryption increases the obstacles that potential attackers must overcome.

[0143] Preferably, the specific steps of distributed storage include:

[0144] The data center performs in-depth processing on the received facial feature set, including feature set splitting, feature set encoding, feature set compression, and feature set encryption. According to the encryption level, the data center configures the replication factor for the original facial image, local features, and global features, and performs redundant backup according to the replication factor; the replication factor determines the number of redundant backups of each data fragment in the distributed storage. For global feature data, a higher replication factor is configured to improve data reliability and fault tolerance.

[0145] A distributed storage architecture is configured for the data center to perform distributed storage of the redundantly backed-up original facial images, local features, and global features. Through load balancing technology, data access requests are dispersed to multiple data nodes in the distributed storage architecture to avoid overloading of a single node.

[0146] Among them, a local storage device is set up inside the vehicle to cache the normalized face images obtained by the edge computing node and the global features extracted by the data center, so as to support basic face recognition functions in the absence of network connection;

[0147] At the same time, configure the update threshold , Used to measure the time interval required for updating. Used to measure the memory capacity that needs to be updated. When the interval between local storage device cache updates is greater than When the local storage device has less memory capacity than When , the normalized face images cached locally are adaptively synchronized to the data center.

[0148] Step S5: The personalized processing unit compares the current driver feature set with the feature set stored in the data center to determine whether the driver's facial image matches to determine the control authority of the vehicle.

[0149] In this example, see Figure 4 , the personalized processing unit receives the current driver's facial image and authenticates the driver through the feature set. If the identity authentication is successful, the system determines the vehicle's control authority, such as automatically adjusting the seat position, playing the driver's favorite music, etc. At the same time, the system continuously monitors the driver's status, including fatigue and distraction. By analyzing changes in facial features, such as the frequency and duration of eye closure, the system can detect whether the driver is fatigued or distracted, and issue an alarm in time to remind the driver to rest or concentrate to ensure driving safety. Accurate verification of the driver's identity, provision of vehicle control authority, and real-time monitoring of driving safety are achieved.

[0150] Preferably, the specific steps of identity verification include:

[0151] Detect whether the current vehicle network is connected. If the network is connected, send a request to the remote data center to obtain a real-time updated facial image feature set, perform complex feature matching on the captured driver's facial image features with the feature set provided by the data center, and determine whether the current driver already exists in the data center of the vehicle based on the matching results.

[0152] The complex feature matching includes:

[0153] By sending a request to a remote data center, a real-time updated facial image feature set is obtained, including facial images corresponding to different facial image identifiers stored in the data center and their local features and global features, etc., for matching with the captured driver facial image;

[0154] Using edge computing nodes, the facial image of the current driver is captured and normalized, and then transmitted to the data center for feature extraction to match the feature set provided by the data center.

[0155] Similarity measurement methods including cosine similarity and Euclidean distance are used to match the captured driver's facial image features with the features provided by the data center. By comparing the similarities between different features, it is possible to more accurately determine whether the current driver already exists in the data center of the vehicle.

[0156] According to the matching result, the identity of the current driver is determined. If the match is successful, it is considered that the current driver already exists in the data center of the vehicle; otherwise, it is considered that the current driver does not exist in the data center.

[0157] According to the result of judging the identity of the driver, corresponding operations are performed, that is, if the current driver already exists in the data center, he can be provided with control authority of the vehicle; otherwise, his access request can be rejected or he can be prompted to try again.

[0158] If the network is not connected, it will enter local verification mode, read the pre-cached normalized facial image and global features from the vehicle's local storage device, use local data for simple feature matching to determine the driver's identity, and determine whether the current driver already exists in the vehicle's local storage device based on the local matching results.

[0159] The simple feature matching includes:

[0160] Obtain the normalized face image of the current driver through the edge computing node, measure the similarity with the normalized face image in the local storage device, configure the similarity threshold, and output the face image identifier whose similarity measurement is greater than the similarity threshold;

[0161] When the number of output facial image identifiers is 0: the current driver's access request is rejected and the vehicle control authority is not provided; the driver is prompted to try again or contact the administrator for processing, and the access attempt is recorded, including the time, location and the driver's facial image information.

[0162] When the number of output facial image identifiers is 1: the recognition result is displayed through a visual interface, and the current driver is requested to confirm. When the driver confirms the recognition result, the control authority of the vehicle corresponding to the current facial image identifier is provided, including adjusting the seat position and playing the driver's favorite music; when the driver fails the recognition result, the driver is prompted to try again or contact the administrator for processing, and the access attempt is recorded, including the time, place and the driver's facial image information.

[0163] When the number of output facial image identifiers is greater than 1: for each output facial image identifier, obtain its corresponding global features, call the edge computing node to obtain the global features of the current driver's facial image, perform similarity measurement, output the facial image identifier with the highest similarity, and request the current driver to confirm. When the driver confirms the recognition result, provide control authority for the vehicle corresponding to the current facial image identifier; when the driver fails the recognition result, prompt the driver to try again or contact the administrator for processing, and record the access attempt, including the time, location and driver's facial image information.

[0164] Preferably, the specific steps of driver status monitoring include:

[0165] According to the distributed storage, local features of the current driver's facial image are obtained, including the position and shape of the eyes, nose and mouth, for driver status analysis and comparison;

[0166] Calculate the driver's eye closure frequency and duration, configure the closure frequency threshold and closure time threshold, and when the eye closure frequency and duration exceed the threshold, the driver is judged to be in a fatigue state;

[0167] If the driver is in a fatigued state, the driver's line of sight and gaze point are monitored to determine whether the driver is distracted. If the driver does not look forward for a long time but looks at other areas, it is judged that the driver is distracted;

[0168] The vehicle provides feedback to the driver through a visual panel. When fatigue or distraction is detected, the vehicle can remind the driver through sound alarms, seat vibrations, or flashing lights in the car. At the same time, warning information is displayed on the instrument panel or central control screen to inform the driver of the current risk level and recommended measures.

[0169] Working principle and effect:

[0170] The distributed storage method of facial image data uses the front-end acquisition device to capture the driver's facial image in real time, and uses the edge computing node to perform dynamic normalization processing, including graying, scaling and cropping. After receiving this data, the data center extracts features, including local features and global features, and builds a feature set. Based on these feature sets, the data center performs splitting, encoding, compression and encryption, and then performs redundant backup and index creation through a distributed storage architecture to ensure data security. The personalized processing unit compares the current driver's feature set with the feature set stored in the data center to achieve identity authentication. Once the verification is successful, the system provides vehicle control authority, such as adjusting the seat position and playing music, and monitors the driver's status.

[0171] It improves the accuracy of identity authentication, enhances data security, provides vehicle control authority, optimizes system performance, adapts to different lighting environments, realizes real-time update and synchronization, and has flexible network connection modes. It improves the accuracy of identity authentication by combining local and global features and using complex similarity measurement methods. Multi-level encryption protects user privacy and data security, while vehicle control authority improves driving experience. Load balancing and local cache strategies optimize system performance, and dynamic normalization processing adapts to different lighting environments. Real-time update and synchronization mechanisms ensure the timeliness of data, while flexible network connection modes ensure system reliability and availability. Multi-level feature extraction comprehensively describes face images and improves recognition accuracy and robustness. Multi-scale LBP feature vectors and HOG feature vectors capture global texture information and edge shape features, respectively. The key feature vector generated during the encryption process ensures encryption security and traceability. The distributed storage architecture improves data reliability and accessibility, while the visual interface enhances user experience and interactivity.

[0172] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A distributed storage method for face image data, characterized in that: include Step S1: Capturing the driver's face image in real time through the front-end acquisition device; Step S2: using the edge computing node to determine the lighting conditions of the captured face image, dynamically normalize the face image according to the lighting conditions, adaptively grayscale the face image based on the lighting conditions, and upload the processed face image to the data center after scaling and cropping; Step S3: the data center performs feature extraction on the face image data after dynamic normalization processing, wherein the feature extraction includes local feature extraction and global feature extraction, and constructs a feature set; Step S4: Based on the feature set, preliminary processing is performed, including feature set splitting, encoding, compression, encryption, and then redundant backup is created and indexing is established in the data center through distributed storage; Step S5: The personalized processing unit compares the current driver feature set with the feature set stored in the data center to determine whether the driver's facial image matches to determine the control authority of the vehicle; The specific steps of the global feature extraction include: Perform global texture extraction on face images, apply neighborhood windows of different sizes to face images, perform multi-scale extraction, calculate the LBP eigenvalue of each window, merge the LBP eigenvalues ​​of all scales, and form a multi-scale LBP feature vector; Perform global edge shape feature extraction on face images, divide the face images into multiple levels of regions, extract HOG features independently on each level of regions, and merge the features of all levels into a long vector; The step of global texture extraction comprises: Configure the neighborhood window size and apply the neighborhood window to each pixel with it as the center; Calculate the LBP feature value of each window, Right now: ,in, is the coordinate of the center pixel, It is The coordinates of the surrounding pixels, Representing coordinates The pixel value at Representing coordinates The pixel value at is the number of pixels in the neighborhood, is a step function, defined as: ,at this time, Pick When Greater than ,but The value is 1, if Less than or equal to ,but The value is 0; Configuring the Offset Angle Threshold , for each pixel, calculate its LBP value within the offset angle threshold range, and take its minimum value as the final LBP feature value of the pixel, that is: ,in, is the rotation angle, and its value range is , is the rotation angle The LBP eigenvalue under ; Merge the LBP eigenvalues ​​of all neighborhood window configurations and merge the LBP eigenvalues ​​obtained at each scale into a long vector.

2. The distributed storage method for facial image data according to claim 1, characterized in that: The specific steps of the dynamic normalization process include: Get a face image and determine the lighting conditions by calculating the statistics of the brightness distribution, contrast, and color saturation of the face image. The calculation formula is as follows: ,in, is the comprehensive evaluation value of lighting conditions. is the mean brightness of the face image, is the standard deviation of face image brightness, is the color saturation of the face image, is the face image contrast, is a positive constant; Configuring light thresholds and ,and Less than ,like , then the face image is judged to be in a low-light environment. If , then the face image is judged to be in a high-light environment, otherwise it is in a normal-light environment.

3. The distributed storage method for facial image data according to claim 1, characterized in that: The specific steps of extracting local features include: Identify key points in the image. For each detected key point, configure a window, divide the pixels in the window into sub-windows, and calculate the gradient histogram of each sub-window. Connect the histograms of all sub-windows to form a descriptor vector. All descriptor vectors are stacked into a matrix, singular value decomposition is performed, and the eigenvectors corresponding to the largest k eigenvalues ​​are selected for data projection to achieve dimensionality reduction; The distance between the descriptor vector to be matched and each descriptor vector in the face database is calculated, and the local facial feature label of the descriptor vector with the smallest distance is used as the label of the descriptor vector to be matched.

4. The distributed storage method for facial image data according to claim 1, characterized in that: The specific steps of extracting the global edge shape feature include: Divide the face image into multiple hierarchical regions to form a multi-level grid structure; For each level of the region, calculate the gradient amplitude and gradient direction of each pixel in the region. For any pixel in the region , the calculation formula of gradient amplitude and gradient direction is: , ,in, It's a pixel The gradient amplitude of It's a pixel The gradient direction of and It's a pixel Horizontal and vertical gradients; The gradient direction is divided into interval directions. For each pixel point, its gradient amplitude is accumulated into the histogram of the corresponding interval according to its gradient direction, and the gradient direction histogram of each area is dynamically normalized. Combine the gradient direction histograms in the regions of all levels into one long vector.

5. The distributed storage method for facial image data according to claim 1, characterized in that: The specific steps of encrypting the feature set include: Based on the encoded and compressed data fragments, different encryption levels are used for the original face image, local features and global features, and multi-level encryption is performed separately, including: The first level of encryption uses an encryption algorithm to perform a first-level encryption on the encoded and compressed original face image data fragment, generates a key and performs key encoding, which is used as a key feature vector and saved in the local feature vector; The second level of encryption uses an encryption algorithm to perform secondary encryption on the local feature data fragments, generates a new key and encodes the key as a key feature vector, which is saved in the global feature vector; The third level of encryption uses an encryption algorithm to perform three-level encryption on the global feature data fragment, generate the final key and encode the key as a key feature vector, and save it in a secure storage area.

6. The distributed storage method for facial image data according to claim 5, characterized in that: The specific steps of the distributed storage include: The data center performs deep processing on the received facial feature set, wherein the deep processing includes feature set splitting, feature set encoding, feature set compression, and feature set encryption; According to the encryption level, respectively configure the replication factor for the original face image, the local features and the global features, and perform redundant backup according to the replication factor; Configure a distributed storage architecture for the data center and perform distributed storage of the redundantly backed-up original face images, local features, and global features; Through load balancing technology, data access requests are distributed to multiple data nodes in the distributed storage architecture to avoid overloading of a single node; A local storage device is set up inside the vehicle to cache the normalized face images obtained by the edge computing node and the global features extracted by the data center; Configuring Update Thresholds , Used to measure the time interval required for updating. Used to measure the memory capacity that needs to be updated. When the interval between local storage device cache updates is greater than When the local storage device has less memory than When , the normalized face images cached locally are adaptively synchronized to the data center.

7. The distributed storage method for facial image data according to claim 1, characterized in that: The specific steps of determining whether the driver's facial image matches include: Detect whether the current vehicle network is connected. If the network is connected, send a request to the remote data center to obtain the real-time updated facial image feature set, perform complex feature matching on the captured driver's facial image features and the feature set provided by the data center, and determine whether the current driver already exists in the data center of the vehicle based on the matching results; If the network is not connected, it will enter local verification mode, read the pre-cached normalized facial image and global features from the vehicle's local storage device, use local data for simple feature matching to determine the driver's identity, and determine whether the current driver already exists in the vehicle's local storage device based on the local matching results.

8. The distributed storage method for facial image data according to claim 7, characterized in that: The simple feature matching includes: Obtain the normalized face image of the current driver through the edge computing node, measure the similarity with the normalized face image in the local storage device, configure the similarity threshold, and output the face image identifier whose similarity measurement is greater than the similarity threshold; When the number of output face image identifiers is 0: the access request of the current driver is rejected and the vehicle control authority is not provided; When the number of output face image identifiers is 1: the recognition result is displayed through a visual interface, and the current driver is requested to confirm; When the number of output facial image identifiers is greater than 1: for each output facial image identifier, obtain its corresponding global features, call the edge computing node to obtain the global features of the current driver's facial image, perform similarity measurement, output the facial image identifier with the highest similarity, and request the current driver to confirm.

Images