Distributed edge calculation face recognition method and system for non-inductive passage

Abstract

The invention relates to the technical field of intelligent recognition, in particular to a distributed edge calculation face recognition method and system for non-inductive passage, and the method comprises the steps: capturing a video stream through edge image collection equipment disposed at a passage entrance, and extracting a face image; and carrying out localized preprocessing, feature extraction and quick comparison with a local cache feature library by the edge computing node. If local matching succeeds, passing is controlled immediately; if the matching fails, the features are uploaded to a central cloud server, enhanced processing and global library comparison are performed by using the stronger computing capability of the central cloud server, and a result is issued and executed, so that the low delay of edge computing and the strong capability of cloud computing are effectively fused, and both the recognition efficiency and the system reliability are considered; and meanwhile, links such as encrypted transmission, dynamic updating, abnormal offline processing and detailed log recording are covered, the security, adaptability and maintainability of the system are remarkably improved while efficient non-inductive traffic is realized, and the method is suitable for intelligent traffic management in a large-scale and high-concurrency scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, in particular to a distributed edge computing face recognition method and system for non-sensing passage. BACKGROUND

[0002] Currently, face recognition technology has become the core means to realize intelligent non-sensing passage. Traditional passage management systems generally use centralized processing architecture, which captures video streams through front-end cameras and transmits all image data containing faces to a remote central server for unified recognition and processing. This mode brings huge computing pressure to the central server when processing large-scale concurrent requests, and high bandwidth occupation and inherent network transmission delay cause the system response time to be prolonged, making it difficult to meet the real-time requirements of non-sensing passage.

[0003] In addition, existing methods face significant privacy and security risks when processing sensitive biometric data. The massive face images are transmitted in plaintext on the network, increasing the possibility of data being stolen or tampered with during transmission. Centralized data storage also makes the central database a major target of attacks, and once leaked, it will cause a serious personal information security crisis. Therefore, how to improve efficiency while ensuring the end-to-end security of biometric data from collection, transmission to processing is an important challenge faced by current technology.

[0004] Finally, the scalability and reliability of existing systems are limited. Their performance is highly dependent on the availability of central cloud servers and network conditions, with a prominent single point of failure risk. When the network connection is unstable or interrupted, the entire recognition process will be stalled, causing passage service interruption. At the same time, system expansion requires constant upgrading of the hardware performance of the central server, which is costly and lacks flexibility, making it difficult to adapt to the demand for elastic computing and high availability in different scale scenarios. SUMMARY

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] According to the first aspect of the present application, a distributed edge computing face recognition method for non-sensing passage is claimed, characterized in that it comprises the following steps:

[0007] S1: Real-time capture of passage area video stream through edge image acquisition devices deployed at the passage entrance, and extraction of the face image to be recognized from the video stream;

[0008] S2: The edge computing node receives the face image to be recognized, calls the locally preloaded lightweight face detection model to perform accurate cropping and alignment of the face area, and obtains a standardized face image;

[0009] S3: The edge computing node invokes a locally deployed first face feature extraction model to process the standardized face image, and generates a corresponding high-dimensional feature vector;

[0010] S4: The edge computing node performs a quick comparison between the high-dimensional feature vector and a locally cached feature vector library, calculates a similarity score, and if the highest similarity score exceeds a first threshold, performs S6; otherwise, performs S5;

[0011] S5: The edge computing node uploads the high-dimensional feature vector to a remote central cloud computing server, the central cloud computing server invokes a deployed second face feature extraction model to perform intensive processing on the received high-dimensional feature vector, generates an intensive feature vector, and compares the intensive feature vector with a central feature vector library, and issues a recognition result and corresponding permission information to the edge computing node;

[0012] S6: The edge computing node generates a pass control instruction according to the local recognition result or the recognition result received from the central cloud, and sends the pass control instruction to a door controller to perform a pass action.

[0013] Further, in step S1, further comprising:

[0014] S1.1: Decoding the pass area video stream frame by frame to obtain a sequence of image frames;

[0015] S1.2: Scanning each image frame using a classifier based on gradient histogram features to preliminarily detect potential face regions;

[0016] S1.3: Positioning key points for each preliminarily detected potential face region, and performing affine transformation based on the key point coordinates to complete face alignment;

[0017] S1.4: Normalizing the size of the aligned face image and performing illumination compensation preprocessing to obtain the face image to be recognized.

[0018] Further, in step S3, further comprising:

[0019] S3.1: Inputting the standardized face image into the first face feature extraction model based on a deep convolutional neural network architecture, and the input layer receiving the standardized face image;

[0020] S3.2: The first face feature extraction model performs hierarchical feature abstraction on the input standardized face image through multiple consecutive convolutional layers and pooling layers, and gradually extracts image features from low-order to high-order;

[0021] S3.3: Before the full connection layer of the convolutional neural network, an embedding layer is introduced to map the abstracted high-order features into an intermediate feature vector of a specified dimension;

[0022] S3.4: The intermediate feature vector is subjected to L2 normalization processing, so that the vector length is 1, and the high-dimensional feature vector is output.

[0023] Further, in step S4, it further includes:

[0024] S4.1: The edge computing node maintains a local feature vector library, which stores the face feature vectors of the subset of recently active users and their corresponding user identifiers;

[0025] S4.2: An approximate nearest neighbor search algorithm is used to calculate the similarity between the high-dimensional feature vector to be compared and all feature vectors in the local library, and the similarity is determined by calculating the cosine distance between two L2 normalized feature vectors;

[0026] S4.3: A dynamic first threshold is set, which is dynamically adjusted by the central cloud computing server according to the security level requirement and network delay condition and is delivered to each edge computing node;

[0027] S4.4: All comparison results higher than the second threshold and less than the first threshold are recorded to form potential matching records for subsequent model optimization and data analysis.

[0028] Further, the update strategy of the locally cached feature vector library is:

[0029] The edge computing node regularly receives feature vector update data packets from the central cloud computing server;

[0030] The update data packet contains the feature vectors of the active user subset dynamically generated according to the user frequency and timestamp information;

[0031] When the storage space of the edge computing node reaches the water line, the least frequently used algorithm is used to replace the least active feature vector in the local library.

[0032] Further, in step S5, it further includes:

[0033] The second face feature extraction model is a deep neural network model with a larger number of parameters than the first face feature extraction model, and the dimension of the enhanced feature vector generated is the same as that of the high-dimensional feature vector.

[0034] Further, the method further includes an exception handling step:

[0035] When the network connection between the edge computing node and the central cloud computing server is interrupted, the edge computing node enters an offline mode;

[0036] In the offline mode, the first threshold is automatically reduced, and only the local cached feature vector library is used for recognition;

[0037] All recognition results and access records will be uploaded in batches to the central cloud computing server for review and archiving after the network is restored.

[0038] Further, the method further comprises a data security step:

[0039] Before uploading the high-dimensional feature vector to the central cloud computing server, the edge computing node uses an encryption algorithm based on an elliptic curve to encrypt the feature vector;

[0040] After the central cloud computing server is processed, the permission information issued is encrypted using the public key of the corresponding edge computing node.

[0041] Further, the method further comprises a log recording step:

[0042] For each recognition event, whether successful or not, the edge computing node generates a structured log record;

[0043] The log record at least includes a timestamp, a device identifier, a digital digest of the captured face image, a comparison result, a used feature library version number, and a finally executed action;

[0044] The log record is periodically compressed and uploaded to the central cloud computing server for constructing an access behavior big data analysis model;

[0045] The central cloud computing server dynamically adjusts the size and content of the active user subset issued to each edge computing node according to the log records uploaded by each edge computing node and the load situation, and uniformly manages the version and update of the first face feature extraction model on all edge computing nodes.

[0046] According to the second aspect of the present application, the present application claims to protect a distributed edge computing face recognition system for non-susceptible access, which comprises:

[0047] One or more processors;

[0048] A memory having one or more programs stored thereon, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the distributed edge computing face recognition method for non-susceptible access.

[0049] This invention relates to the field of intelligent recognition technology, specifically a distributed edge computing face recognition method and system for seamless access control. It captures video streams and extracts facial images using edge image acquisition devices deployed at access points. Edge computing nodes perform localized preprocessing, feature extraction, and rapid comparison with a locally cached feature library. If a local match is successful, access is immediately controlled; if a match fails, the features are uploaded to a central cloud server, utilizing its superior computing power for enhanced processing and global library comparison. The results are then distributed for execution. This effectively integrates the low latency of edge computing with the powerful capabilities of cloud computing, balancing recognition efficiency and system reliability. It also includes encrypted transmission, dynamic updates, offline anomaly handling, and detailed logging. While achieving efficient seamless access control, it significantly improves system security, adaptability, and maintainability, making it suitable for intelligent access management in large-scale, high-concurrency scenarios. Attached Figure Description

[0050] Figure 1 A flowchart illustrating the workflow of a distributed edge computing face recognition method for seamless access claimed in this application embodiment;

[0051] Figure 2 This is a second flowchart of a distributed edge computing face recognition method for seamless access claimed in an embodiment of this application;

[0052] Figure 3 A third flowchart illustrating a distributed edge computing face recognition method for seamless access claimed in this application embodiment;

[0053] Figure 4 The fourth flowchart illustrates a distributed edge computing face recognition method for seamless access claimed in this application. Detailed Implementation

[0054] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0055] The terms "first", "second", "third", etc. in the present application are only for descriptive purpose and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second", "third" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative position relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications also change accordingly. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0056] Reference herein to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment can be included in at least one embodiment of the application. The appearance of the phrase in various places in the specification does not necessarily all refer to the same embodiment, nor is it necessarily mutually exclusive of other embodiments. It is explicitly and implicitly understood that the embodiments described herein can be combined with other embodiments.

[0057] According to the first embodiment of the present application, the present application claims a distributed edge computing face recognition method for non-inductive passing, referring to Figure 1 , comprising the following steps:

[0058] S1: Real-time capture the passing area video stream through the edge image acquisition device deployed at the passing entrance, and extract the to-be-recognized face image therefrom;

[0059] S2: The edge computing node receives the to-be-recognized face image, calls the locally preloaded lightweight face detection model to perform face region accurate cropping and alignment, and obtains a standardized face image;

[0060] S3: The edge computing node calls the locally deployed first face feature extraction model to process the standardized face image, and generates a corresponding high-dimensional feature vector;

[0061] S4: The edge computing node performs fast comparison between the high-dimensional feature vector and the locally cached feature vector library, calculates a similarity score, and if the highest similarity score exceeds a first threshold, executes S6; otherwise, executes S5;

[0062] S5: The edge computing node uploads the high-dimensional feature vector to a remote central cloud computing server, which invokes a deployed second face feature extraction model to perform intensive processing on the received high-dimensional feature vector, generates an intensive feature vector, and compares the intensive feature vector with a central feature vector library, and issues an identification result and corresponding permission information to the edge computing node;

[0063] S6: The edge computing node generates a pass control instruction according to the local identification result or the identification result received from the central cloud, and sends the pass control instruction to a door controller to execute a pass action.

[0064] In this embodiment, in S1, a plurality of high-definition cameras fixedly deployed at a passageway entrance continuously acquire video stream data of the passageway; a primary processing unit integrated in the camera or in the vicinity analyzes the video stream in real time, actively detects and intercepts an original image frame containing a face region.

[0065] In S2, an edge computing node located at the entrance site receives the original image frame through an internal local area network; the node invokes a pre-stored face key point positioning model in an internal memory to outline a contour and mark feature points of the face in the image frame; according to the marked feature point coordinates, the face posture is corrected to a standard front view angle through geometric transformation, and an image block containing only the face region is cropped, to complete standardization processing.

[0066] In S3, the edge computing node inputs the standardized face image block into a neural network model pre-installed in the node; the model is composed of a series of convolution, nonlinear activation and pooling operations, and is used to extract details, local and global features in the image layer by layer; the final output layer of the model generates a fixed-length and numerical feature sequence, i.e., a high-dimensional feature vector.

[0067] In S4, the edge computing node accesses a local database storing a set of face feature vectors of part of authorized users; the node uses a vector similarity measurement method to calculate the matching scores of the to-be-identified feature vector and each stored vector in the library one by one; the system presets a higher confidence threshold, if the highest matching score retrieved exceeds the threshold, S6 is entered; if all matching scores are lower than the threshold, S5 is entered.

[0068] In S5, the edge computing node transmits the to-be-identified feature vector to a remote central cloud computing cluster through a network; the central cloud invokes a more complex and precise large neural network model to perform secondary deep processing on the received vector, to extract more distinctive feature expressions; the processed new vector is compared with a central database storing information of all authorized users, and the obtained user identity and pass permission conclusion are returned to the edge computing node initiating the request.

[0069] S6. The edge computing node comprehensively judges the result: if the recognition is successful and the user has the access right, a "permit access" instruction signal is generated and sent to the electronic door lock or gate controller through the hardware interface; if the recognition fails or the right is insufficient, a "refuse access" instruction is generated or an abnormal processing procedure is triggered.

[0070] Further, with reference to Figure 2 In step S1, further comprising:

[0071] S1.1: decoding the access area video stream frame by frame to obtain a sequence of image frames;

[0072] S1.2: scanning each image frame using a classifier based on gradient histogram features to preliminarily detect potential face regions;

[0073] S1.3: positioning key points for each of the preliminarily detected potential face regions, and performing affine transformation based on the key point coordinates to complete face alignment;

[0074] S1.4: normalizing the size of the aligned face image and performing illumination compensation preprocessing to obtain the face image to be recognized.

[0075] In this embodiment, specifically comprising:

[0076] S1.1: sampling the continuous video stream output by the camera at a fixed frame rate, and decomposing the video data into a series of static digital images arranged in time sequence.

[0077] S1.2: performing grayscale conversion on each frame of digital image, and then using a sliding window detection-based strategy to analyze window regions of different sizes, quickly exclude non-face regions, and preliminarily locate one or more candidate boxes that may contain faces.

[0078] S1.3: for each of the preliminarily detected face candidate boxes, using a regression prediction method to accurately find key pixel points that identify eye, nose, and mouth corner positions.

[0079] S1.4: according to the spatial distribution relationship of these key points, the tilted and side-turned face image is rotated and corrected through coordinate transformation; then the corrected face image is scaled to a uniform pixel size, and the histogram equalization technique is used to eliminate the effects of uneven illumination, and finally a consistent quality of the image to be recognized is output.

[0080] Further, with reference to Figure 3 , in step S3, further comprising:

[0081] S3.1: input the standardized face image into the first face feature extraction model based on a deep convolutional neural network architecture, and an input layer receives the standardized face image;

[0082] S3.2: the first face feature extraction model abstracts the input standardized face image through a plurality of consecutive convolutional layers and pooling layers, and gradually extracts image features from low order to high order;

[0083] S3.3: before the fully connected layer of the convolutional neural network, an embedding layer is introduced to map the abstracted high-order features to an intermediate feature vector of a specified dimension;

[0084] S3.4: L2 normalization is performed on the intermediate feature vector, so that the vector length is 1, and the high-dimensional feature vector is output.

[0085] In this embodiment, it also includes:

[0086] S3.1: the standardized face image obtained in step S2 is input into the first face feature extraction model, which is a deep neural network trained on a large amount of face image data.

[0087] S3.2: the neural network includes a plurality of processing stages, each stage being composed of convolution operation, activation function and pooling operation; the convolution operation uses filters of different sizes to slide on the image to extract basic visual elements such as edges and textures; the activation function introduces a nonlinear transformation to the calculation result; the pooling operation reduces the sampling of the feature map, compresses the data volume and enhances the translation invariance of the features.

[0088] S3.3: after the above-mentioned alternating processing of multiple stages, the network expands and converts the obtained two-dimensional feature map into a one-dimensional feature sequence.

[0089] S3.4: amplitude normalization calculation is performed on the one-dimensional feature sequence, i.e. each numerical element in the sequence is divided by the Euclidean norm of the sequence, so that the high-dimensional feature vector output finally is on a hypersphere with a standard length, and this operation is to optimize the accuracy and stability of subsequent vector similarity comparison.

[0090] Further, with reference to Figure 4 , in step S4, it also includes:

[0091] S4.1: the edge computing node maintains a local feature vector library, which stores face feature vectors of a subset of recent active users and their corresponding user identifiers;

[0092] S4.2: Using the approximate nearest neighbor search algorithm, the high-dimensional feature vector to be compared is calculated with all feature vectors in the local library, and the similarity is determined by calculating the cosine distance between the two L2 normalized feature vectors;

[0093] S4.3: Set a dynamic first threshold, which is dynamically adjusted by the central cloud computing server according to the security level requirement and network delay condition and is distributed to each edge computing node;

[0094] S4.4: Record all comparison results higher than the second threshold and less than the first threshold to form potential matching records for subsequent model optimization and data analysis.

[0095] In this embodiment, it also includes:

[0096] S4.1: The edge computing node maintains a feature vector cache library in the local memory or high-speed storage device, which is not a full-amount database, but a high-frequency user feature subset dynamically selected by the central cloud system according to the user's recent access history; Each feature vector is associated with the user's unique ID and necessary permission information.

[0097] S4.2: In the comparison process, the cosine similarity between the feature vector to be identified and each vector in the cache library is calculated; The calculation is a mathematical process of multiplying the corresponding elements of two vectors and summing them up, and the result value range is clear, and the closer to 1 indicates the higher similarity.

[0098] S4.3: The first threshold is a dynamically configurable floating point number, whose value is calculated and distributed by the central cloud management system according to the global security policy and real-time network condition, and the edge computing node receives and applies the latest threshold.

[0099] S4.4: The system also sets a slightly lower second threshold to record those comparison attempts that do not pass clearly but have certain similarity; These "potential matching" records will be marked and temporarily stored for subsequent audit tracking, or as difficult samples for focused learning in model retraining.

[0100] Further, the update strategy of the local cached feature vector library is:

[0101] The edge computing node regularly receives feature vector update data packets from the central cloud computing server;

[0102] The update data packet contains the feature vectors of the active user subset dynamically generated according to the user access frequency and timestamp information;

[0103] When the storage space of the edge computing node reaches the waterline, the least recently used algorithm is used to replace the least active feature vector in the local library.

[0104] In this embodiment, the central cloud computing server continuously monitors the access records of each user and counts the access frequency of each user in different time periods.

[0105] The central cloud predicts the list of users most likely to access in the near future based on the statistical results and packs the feature vectors of these users into an update file.

[0106] The update file is periodically pushed over the network or pulled by the edge computing node to each edge computing node.

[0107] After the edge computing node receives the update file, it merges the feature vector data in it into the local cache library.

[0108] When the storage capacity of the local cache library is about to be exhausted, the system starts the elimination mechanism to remove the user feature vectors that have been compared the least number of times or have the earliest update time in the recent past, to make room for new data.

[0109] Further, step S5 further includes:

[0110] The second face feature extraction model is a deep neural network model with a larger number of parameters than the first face feature extraction model, and the dimension of the enhanced feature vector generated is the same as that of the high-dimensional feature vector.

[0111] Further, the method further includes an exception handling step:

[0112] When the network connection between the edge computing node and the central cloud computing server is interrupted, the edge computing node enters offline mode;

[0113] In offline mode, the first threshold is automatically lowered, and only the feature vector library cached locally is used for recognition;

[0114] All recognition results and access records will be uploaded to the central cloud computing server in bulk for review and archiving after the network is restored.

[0115] Further, the method further includes a data security step:

[0116] Before uploading the high-dimensional feature vector to the central cloud computing server, the edge computing node uses an encryption algorithm based on elliptic curves to encrypt the feature vector;

[0117] After the central cloud computing server finishes processing, the permission information issued is encrypted using the public key of the corresponding edge computing node.

[0118] In this embodiment, the edge computing node is built-in with a heartbeat mechanism, which continuously detects the network connection status with the central cloud server; once a connection timeout or interruption is detected, the offline mode switching is triggered immediately.

[0119] In offline mode, the system automatically lowers the confidence threshold (first threshold) required for successful identification, to relax the standard of local identification, to adapt to the situation where the central cloud cannot be consulted.

[0120] All the access events that occur during the offline period, the process data, including the captured images, feature vectors, comparison results and executed actions, are recorded in the non-volatile memory of the edge computing node.

[0121] When the network connection is restored, the edge computing node automatically uploads the temporarily stored offline records in chronological order to the central cloud server in batches; the central cloud server uses its full-amount database to review these records after the fact, verify the accuracy of the edge recognition results, and update the final state of the access log.

[0122] Further, the method further comprises a log recording step:

[0123] The edge computing node generates a structured log record for each identification event, whether successful or not;

[0124] The log record contains at least a timestamp, device identification, digital digest of the captured face image, comparison result, version number of the used feature library, and the final executed action;

[0125] The log record is periodically compressed and uploaded to the central cloud computing server for building an access behavior big data analysis model;

[0126] The central cloud computing server dynamically adjusts the size and content of the active user subset issued to each edge computing node according to the log records uploaded by each edge computing node and the load situation, and uniformly manages the version and update of the first face feature extraction model on all edge computing nodes.

[0127] In this embodiment, the edge computing node requests encryption services from its built-in security password module before preparing to upload the feature vectors.

[0128] The password module uses the elliptic curve public key obtained from the central certificate authority to perform encryption operations on the feature vector data, generating a piece of ciphertext.

[0129] The edge computing node transmits the ciphertext rather than the plaintext feature vector to the central cloud computing server.

[0130] The central cloud server decrypts the received ciphertext using the corresponding private key it holds, restores the original feature vector data, and processes it.

[0131] When issuing instructions, the central cloud server also queries the public key of the target edge computing node, encrypts the instruction information, and ensures that only the target edge computing node can decrypt and read it using its private key.

[0132] For each identification attempt, whether successful or not, the edge computing node automatically generates a record.

[0133] Each record contains precise time information, the device number of the node, the hash check value of the captured face image, all major results compared with the local library, interaction records requested from the central cloud, the final decision, and the feedback state of the controlled hardware device.

[0134] These log records are temporarily stored in a structured data format locally.

[0135] The edge computing node initiates a log upload task at a preset period or when the log file reaches a certain size.

[0136] The accumulated log files are compressed and encrypted and then transmitted to the central cloud's designated log collection server.

[0137] The central cloud server receives and decompresses these logs and stores them in the big data analysis platform for cluster behavior analysis, system performance evaluation, and model optimization iteration.

[0138] The central cloud server analyzes the logs of each edge computing node and identifies high-frequency pass users in the location and time period of the node.

[0139] Based on the analysis results, the central cloud generates a customized "active user subset" recommendation list for each edge computing node in the next update cycle.

[0140] At the same time, the central cloud monitors the computing load and storage usage of each edge computing node. For nodes with lighter loads, larger cache library capacities can be allocated; for nodes with heavier loads, the subset can be simplified to ensure identification speed.

[0141] For the first face feature extraction model, the central cloud maintains version information for all models. When a model needs to be upgraded, the central cloud pushes new model files to edge computing nodes in batches and instructs the nodes to complete model switching and loading during business off-peak periods to ensure system service continuity.

[0142] According to the second embodiment of the present application, the present application claims a distributed edge computing face recognition system for non-sensing pass, which comprises:

[0143] one or more processors;

[0144] a memory having stored thereon one or more programs that, when executed by the one or more processors, cause the one or more processors to implement the distributed edge computing face recognition method for non-susceptible passing.

[0145] The system is deployed at the main entrance of a large enterprise park, and includes the following physical components: multiple high-definition network cameras deployed above the passageway, industrial-grade edge computing gateways installed beside each gate, a central cloud computing server cluster located in the data center room, and controlled electronic gate devices.

[0146] The specific implementation process is as follows:

[0147] In the image capture stage, the network camera continuously captures video data of the passing area. The intelligent analysis unit built-in the camera monitors the video stream in real time, and automatically triggers the face detection program when detecting that a moving object enters the preset monitoring area. The analysis unit adopts a multi-scale detection strategy, which can effectively locate the face area at different distances and angles, and automatically select the frame with the best image quality for output. The output face image is converted in format after preliminary conversion, and is transmitted to the designated edge computing gateway through a special network.

[0148] In the edge preprocessing stage, after the edge computing gateway receives the image data, it starts the preprocessing process. First, the image is converted to a grayscale image to reduce processing complexity. Then, a preloaded face key point positioning model is called, which can accurately identify the positions of multiple feature points on the face, including the exact positions of the eye corners, the coordinate points of the nose tip, and the key points of the lip contour. Based on the spatial distribution relationship of these feature points, the system calculates the transformation parameters of the face pose, and corrects the face image to the standard front view angle through geometric transformation methods in digital image processing. Finally, the corrected face area is standardized and cropped to generate the recognized image that meets the specification requirements.

[0149] In the feature extraction stage, the preprocessed standard face image is sent to the feature extraction module. This module adopts a deep neural network architecture, which is composed of multiple feature extraction layers. The initial layer is responsible for extracting basic visual features such as edges and texture information; the middle layer gradually combines these basic features to form more complex facial feature representations; and the final layer maps these high-level features to a fixed-dimensional feature space to produce a high-dimensional vector representing the face features. The entire processing process is completed locally on the edge gateway without relying on external computing resources.

[0150] The generated feature vector is then sent to the local matching module. This module accesses a feature database stored in the gateway's local memory, which is a subset of active user features dynamically issued by the central server based on recent access conditions. The matching process uses vector similarity calculation methods to compare the input feature vector with all stored vectors in the database, and calculates the similarity score. The system has a dynamically adjustable similarity threshold, and when the highest similarity score exceeds the set threshold, the system immediately completes identity verification.

[0151] When the local matching fails to find a sufficiently similar record, the edge gateway initiates a cloud collaboration process. First, the feature vector to be identified is securely packaged and transmitted to the central cloud computing platform through an encrypted channel. After receiving the request, the central cloud uses a more complex feature processing model to deepen the processing of the vector, enhancing its feature expression ability, and then performs a comprehensive search in the complete personnel feature database. Regardless of whether the identification is successful, the central cloud will generate detailed response information and return it to the edge gateway through a secure channel.

[0152] In the execution control phase, the edge gateway generates control instructions based on the final identification result. For authorized personnel whose identification is successful, the gateway sends an opening signal to the gate controller through a digital interface, and records the relevant information of the access event. For cases where identification fails or is not authorized, the gateway keeps the gate closed and can optionally trigger an alarm device. All operation results are recorded in detail and synchronized to the central management system.

[0153] Test verification data, the system has undergone long-term operation test, the results show that in the normal working period, most of the identification requests can be completed in the edge layer, the average response time remains at a very low level. A small number of requests requiring cloud assistance are also completed within a reasonable time, the overall system identification accuracy remains at a high level, fully meeting the application requirements of non-contact access.

[0154] During the image acquisition process, the network camera captures video streams using the progressive scan method, and each frame of image is adjusted for automatic exposure and white balance to ensure stable image quality. Advanced compression encoding format is used for video stream transmission, which effectively reduces the bandwidth demand while ensuring image quality.

[0155] During face detection, the intelligent analysis unit monitors the motion targets in the video stream in real time. When a moving target with human characteristics is detected, the face detection program is automatically started. The detection process uses a multi-stage screening strategy, first quickly excludes obvious non-face image areas, then performs fine analysis on the candidate areas, and finally determines the accurate face bounding box.

[0156] The detected face region in the image optimization process is subjected to a series of image enhancement processes. These include brightness equalization to eliminate the effects of uneven lighting, adaptive contrast adjustment to enhance facial feature clarity, and image sharpening to improve feature recognizability. These processes significantly improve the accuracy of the subsequent recognition stage.

[0157] For quality evaluation mechanism, the system integrates an image quality evaluation module to conduct multi-dimensional quality evaluation on the captured face image, including clarity score, pose angle evaluation, and lighting condition evaluation. Only images that meet the quality standards will enter the subsequent processing stage, ensuring the reliability of the recognition process.

[0158] For edge feature vector generation, after the standardized face image is input into the neural network during the forward propagation process, it first undergoes convolution operation to extract primary visual features. These features are transformed by a nonlinear activation function and then input into the pooling processing layer for feature dimension reduction. Multiple such combination layers are stacked in sequence to gradually abstract higher-level feature representations.

[0159] During feature mapping and transformation, the network deep layer adopts a special feature mapping structure to convert spatial features into high-dimensional vector representations. This conversion process preserves the discriminative information of face features while eliminating irrelevant variation factors such as lighting changes and expression differences.

[0160] For vector normalization processing, the generated feature vectors undergo strict normalization processing to unify the vector length. This process significantly improves the stability of vector similarity calculation and enhances the accuracy of the subsequent matching stage.

[0161] Computational optimization includes multiple optimizations for neural network computation by the edge gateway, including computational graph optimization, memory reuse strategy, and instruction-level parallel processing, ensuring the efficiency of the feature extraction process.

[0162] The local feature library maintained by the edge gateway adopts an efficient data organization structure to support fast similarity search. Each feature vector is associated with complete identity information and security permission data, which are stored in encrypted form to ensure data security.

[0163] The matching process uses an optimized vector similarity calculation method to calculate the similarity scores between the feature to be identified and all features in the library. The calculation process fully utilizes hardware acceleration capabilities to achieve extremely fast matching.

[0164] The system dynamically adjusts the similarity threshold value based on security requirements and operating environment. It automatically increases the threshold value during periods of high security requirements and maintains the standard threshold value during normal operating conditions, achieving the best balance between security and efficiency.

[0165] All matching attempts, whether successful or not, are recorded in detail. Records of near-matches that do not succeed are marked for special analysis and system optimization.

[0166] The central server intelligently predicts feature library update needs based on actual traffic patterns at each gate. The update strategy takes into account various factors, including personnel traffic frequency, time period characteristics, and special security needs.

[0167] Feature data transmission uses a differential update mechanism, transmitting only changed data, significantly reducing network bandwidth requirements. The transmission process enables strong encryption protection to ensure data security.

[0168] The edge gateway uses an intelligent cache management algorithm to dynamically adjust storage content based on feature usage frequency, ensuring that the most commonly used features remain in the local library.

[0169] The edge gateway implements continuous network connectivity detection, confirming connection status with the central server through regular heartbeat packet exchange. When network anomalies are detected, the failover mechanism is automatically triggered.

[0170] In the event of network interruption, the system automatically switches to offline operation mode. At this time, the recognition strategy is adjusted, and the recognition standard is appropriately relaxed to ensure that basic traffic functions are not affected.

[0171] After network recovery, the system automatically starts the data synchronization process, uploading traffic records and security events accumulated during the offline period to the central server, ensuring data integrity and consistency.

[0172] All sensitive data is strongly encrypted before transmission, using industry-standard encryption algorithms and key management schemes to ensure confidentiality and integrity during data transmission.

[0173] A two-way identity authentication mechanism is established between the edge gateway and the central server to prevent unauthorized access and man-in-the-middle attacks.

[0174] The system records all security-related events, including encryption operations, key usage, and access attempts, providing complete data support for security audits.

[0175] The system automatically generates detailed operation logs at each key processing step, including timestamp, device identification, processing results, and performance indicators.

[0176] After standardized processing, log data is compressed and encrypted for storage, and is regularly transmitted to the central analysis platform.

[0177] The central platform conducts in-depth analysis of log data, generating operation reports, performance statistics, and security analysis results to provide data support for system optimization and operational decision-making.

[0178] The central cloud platform monitors the running state and resource usage of each edge computing node in real time, and dynamically adjusts the resource allocation strategy according to the actual load.

[0179] The system supports centralized configuration management, and can batch issue configuration updates and policy adjustments to ensure consistency and timeliness of the entire system configuration.

[0180] Based on the analysis results of the running data, the central platform continuously optimizes system parameters and algorithm strategies to continuously improve the overall performance and user experience of the system.

[0181] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0182] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit. The above is only an embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

[0183] The specific embodiments of the application are described in detail above, but they are only examples. The present application is not limited to the above-described specific embodiments. Any equivalent modification or substitution of the present application made by those skilled in the art is also within the scope of the present application, and therefore, any equivalent transformation and modification, improvement, etc. made without departing from the spirit and principle range of the present application should be covered within the scope of the present application.

Claims

1. A distributed edge computing face recognition method for non-inductive passing, characterized in that, The method comprises the following steps: S1: capturing a video stream of a passing area in real time by an edge image acquisition device deployed at the edge of a passing entrance, and extracting a to-be-identified face image therefrom; S2: an edge computing node receives the to-be-identified face image, calls a locally preloaded lightweight face detection model to perform accurate cropping and alignment of a face region, and obtains a standardized face image; S3: the edge computing node calls a locally deployed first face feature extraction model to process the standardized face image, and generates a corresponding high-dimensional feature vector; S4: the edge computing node performs rapid comparison between the high-dimensional feature vector and a locally cached feature vector library, calculates a similarity score, and if the highest similarity score exceeds a first threshold, executes S6; otherwise, executes S5; S5: the edge computing node uploads the high-dimensional feature vector to a remote central cloud computing server, the central cloud computing server calls a deployed second face feature extraction model to perform intensive processing on the received high-dimensional feature vector, generates an intensive feature vector, and performs comparison between the intensive feature vector and a central feature vector library, and issues an identification result and corresponding permission information to the edge computing node; S6: the edge computing node generates a passing control instruction according to a local identification result or an identification result received from the central cloud, and sends the passing control instruction to a door access controller to execute a passing action.

2. The face recognition method of claim 1, wherein, In step S1, further comprising: S1.1: decoding the passing area video stream frame by frame to obtain sequential image frames; S1.2: scanning each image frame using a classifier based on gradient histogram features to preliminarily detect potential face regions; S1.3: performing key point positioning on each of the preliminarily detected potential face regions, and performing affine transformation according to the key point coordinates to complete face alignment; S1.4: normalizing the size of the aligned face image, and performing preprocessing such as illumination compensation to obtain the to-be-identified face image.

3. The face recognition method of claim 1, wherein, In step S3, further comprising: S3.1: inputting the standardized face image into the first face feature extraction model based on a deep convolutional neural network architecture, and an input layer receiving the standardized face image; S3.2: the first face feature extraction model performs hierarchical feature abstraction on the input standardized face image through multiple consecutive convolutional layers and pooling layers, and gradually extracts image features from low-order to high-order; S3.3: before the fully connected layer of the convolutional neural network, an embedding layer is introduced to map the abstracted high-order features to an intermediate feature vector of a specified dimension; S3.4: performing L2 normalization processing on the intermediate feature vector to make the vector length 1, and outputting the high-dimensional feature vector.

4. The face recognition method of claim 1, wherein, In step S4, further comprising: S4.1: the edge computing node maintains a local feature vector library, and the local feature vector library stores face feature vectors and corresponding user identifiers of a subset of recent active users; S4.2: Using the approximate nearest neighbor search algorithm, the high-dimensional feature vector to be compared is calculated with all the feature vectors in the local library, and the similarity is determined by calculating the cosine distance between the two L2 normalized feature vectors; S4.3: Set a dynamic first threshold, which is dynamically adjusted by the central cloud computing server according to the security level requirement and network delay condition and is delivered to each edge computing node; S4.4: Record all the comparison results higher than the second threshold and less than the first threshold to form potential matching records for subsequent model optimization and data analysis.

5. The face recognition method of claim 1, wherein, The updating strategy of the locally cached feature vector library is: The edge computing node regularly receives feature vector update data packets from the central cloud computing server; The update data packet contains the feature vectors of the active user subset dynamically generated according to the user frequency and timestamp information; When the storage space of the edge computing node reaches the water line, the least frequently used algorithm is used to replace the least active feature vector in the local library.

6. The face recognition method of claim 1, wherein, In step S5, it also includes: The second face feature extraction model is a deep neural network model with a larger number of parameters than the first face feature extraction model, and the dimension of the generated enhanced feature vector is the same as that of the high-dimensional feature vector.

7. The face recognition method of claim 1, wherein, The method further includes an abnormality handling step: When the network connection between the edge computing node and the central cloud computing server is interrupted, the edge computing node enters offline mode; In offline mode, the first threshold is automatically lowered, and recognition is performed only based on the locally cached feature vector library; All recognition results and access records will be uploaded to the central cloud computing server in batches after the network is restored for review and archiving.

8. The face recognition method of claim 1, wherein, The method further includes a data security step: Before uploading the high-dimensional feature vector to the central cloud computing server, the edge computing node uses an encryption algorithm based on elliptic curve to encrypt the feature vector; After the central cloud computing server finishes processing, the permission information delivered is encrypted using the public key of the corresponding edge computing node.

9. The face recognition method of claim 1, wherein, The method further includes a log recording step: For each recognition event, whether successful or not, the edge computing node generates a structured log record; The log record at least includes timestamp, device identification, digital digest of the captured face image, comparison result, feature library version number used, and final executed action; The log record is periodically compressed and uploaded to the central cloud computing server for building access behavior big data analysis model; The central cloud computing server dynamically adjusts the size and content of the active user subset delivered to each edge computing node according to the log records and load conditions uploaded by each edge computing node, and uniformly manages the version and update of the first face feature extraction model on all edge computing nodes.

10. A distributed edge computing face recognition system for non-inductive passing, characterized in that, It includes: One or more processors; A memory having one or more programs stored thereon, when the one or more programs are executed by the one or more processors, the one or more processors implement a distributed edge computing face recognition method for non-susceptible access according to any one of claims 1 to 9.

Images