Rail transit access control method and read-write industrial control all-in-one machine adopting the same

Abstract

The application relates to a rail transit passing control method and a read-write industrial control all-in-one machine adopting the method, and the method comprises the following steps: collecting passenger biological characteristics or electronic certificates through a multi-modal recognition device, and associating environment and passenger flow data. When verification is performed, dynamic adaptation is performed: the biological characteristics are matched according to scene parameter adjustment; the electronic certificate is checked for state, space-time compliance and cross-verification with the bound biological characteristics. Finally, through a three-level weight algorithm of a characteristic layer, an evidence layer and a decision layer, verification results, scene labels, passenger flow density and passenger passing posture data are fused to generate a comprehensive judgment value to control the opening and closing of a gate, and a scene-based prompt is output when the verification fails. The application has the following effects: the passing verification is upgraded from single static verification to multi-dimensional dynamic intelligent decision-making which fuses identity, behavior, environment and state, and under the premise of ensuring safety, the passing efficiency and passenger convenience are systematically improved.

Description

Technical Field

[0001] This invention relates to the field of rail transit passenger identification technology, and in particular to a rail transit traffic control method and an integrated industrial control machine for reading and writing using this method. Background Technology

[0002] In the field of modern rail transit, with the continuous advancement of technology, traffic control methods are developing towards intelligence and diversification. The accurate matching of identity authentication and traffic rights is of great significance for improving traffic efficiency and protecting the rights and interests of passengers, especially special groups.

[0003] Currently, rail transit access control systems typically use turnstiles to identify and control passenger passage. The turnstiles identify relevant passenger characteristics and then authenticate them, such as through facial recognition or electronic credentials. This process usually involves a specific identification method to verify passenger information and compare it with pre-stored information in a pre-configured database. If the comparison is successful, the passenger is allowed to pass.

[0004] Existing technologies have significant shortcomings in dynamic adaptability and accurate matching, specifically as follows: relying on a single identification method, once the feature cannot be effectively collected due to objective conditions, there is a lack of alternative verification paths, directly blocking passage, which not only reduces passage efficiency but also increases security risks. Summary of the Invention

[0005] In order to upgrade access verification from a single static verification to a multi-dimensional dynamic intelligent decision-making that integrates identity, behavior, environment, and status, and to systematically improve access efficiency and passenger convenience while ensuring security, this application provides a rail transit access control method and an integrated reading and writing industrial control machine using the method.

[0006] Firstly, this application provides a method for controlling rail transit traffic, which adopts the following technical solution:

[0007] A method for controlling rail transit traffic includes:

[0008] The system collects and acquires passengers' biometric or electronic credential information, with the biometrics being collected and acquired through a multimodal identity recognition device deployed on a read / write industrial control all-in-one machine.

[0009] The collected biometric data or electronic voucher information is converted into identity information in a unified format and synchronously linked to environmental parameters and passenger flow density at the time of collection.

[0010] Dynamic adaptation verification is performed on identity information. Scene labels are generated based on environmental parameters. If the collected information is biometric information, feature vectors are extracted from the feature extraction parameter set called from the preset parameter library based on the scene label, and matched with the sub-templates of the corresponding scene in the feature template library to output the biometric verification result. If the collected information is electronic voucher information, the voucher identifier is extracted, the ticket status is determined based on the ticket's full life cycle data, and the spatiotemporal compliance is determined by combining the terminal location and time period. The current and bound historical biometric features are cross-verified to determine the degree of consistency, and the verification result of the fusion of ticket status, spatiotemporal compliance and degree of consistency is output.

[0011] The verification results are integrated, combining scene labels, passenger flow density in the passageway, and passenger walking posture data perceived by the infrared sensor array. A three-level weighted algorithm of feature layer, evidence layer, and decision layer is used to generate a comprehensive judgment value, which is then compared with a preset threshold to generate a comprehensive judgment result. Specifically, the feature layer assigns confidence weights to the verification results based on scene labels; the evidence layer calculates the evidence support of the verification results based on the weights and preset rules; and the decision layer determines the anomaly index based on the walking posture data, determines the correction coefficient based on the passenger flow density in the passageway, and integrates the evidence support, anomaly index, and correction coefficient to generate a comprehensive judgment value.

[0012] If the overall judgment result is "pass", the gate will be opened and the encrypted transaction log will be recorded.

[0013] If the overall judgment result is "not passed", a scenario-based prompt will be output.

[0014] By adopting the above technical solution, this method integrates multimodal verification and real-time environmental perception, and significantly improves identity recognition accuracy and anti-counterfeiting capabilities through dynamic adaptation and a three-level weighting algorithm. Combining passenger flow density and traffic posture optimization decisions, it improves traffic efficiency while ensuring safety, achieving scenario-based intelligent management and control.

[0015] Secondly, this application proposes a read / write industrial control all-in-one machine, which adopts the following technical solution:

[0016] A read / write industrial control all-in-one machine, comprising:

[0017] The data collection unit is used to collect and obtain passengers' biometric or electronic credential information.

[0018] The conversion unit is used to convert the collected biometric data or electronic voucher information into identity information in a unified format, and synchronously associate it with environmental parameters and passenger flow density at the time of collection.

[0019] The verification unit is used to perform dynamic adaptation verification on identity information. It generates scene labels based on environmental parameters. If the collected information is biometric information, it extracts feature vectors from the feature extraction parameter set called from the preset parameter library based on the scene label, and matches them with the sub-templates of the corresponding scene in the feature template library, outputting the biometric verification result. If the collected information is electronic voucher information, it extracts the voucher identifier, determines the ticket status based on the ticket's full life cycle data, determines spatiotemporal compliance by combining the terminal location and time period, triggers cross-verification of current and bound historical biometrics to determine the degree of consistency, and outputs the verification result that integrates the ticket status, spatiotemporal compliance, and degree of consistency.

[0020] The fusion decision unit is used to fuse verification results. It combines scene labels, channel passenger flow density and passenger passage posture data perceived by infrared sensor arrays, and generates a comprehensive judgment value through a three-level weight algorithm of feature layer-evidence layer-decision layer. It then compares the comprehensive judgment result with a preset threshold.

[0021] The execution unit is used to drive the gate to open and record the encrypted transaction log when the comprehensive judgment result is "pass", and to output a scenario-based prompt when the comprehensive judgment result is "fail". Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of a rail transit traffic control method according to an embodiment of this application. Detailed Implementation

[0023] Reference Figure 1 This application discloses a method for controlling traffic flow in rail transit. It includes the following steps:

[0024] Step S1 involves collecting the passenger's biometric or electronic credential information. Biometrics are collected using a multimodal identity recognition device deployed on a read / write industrial control computer. In this step, biometrics refer to facial features and palm vein features; electronic credentials refer to QR codes, NFC cards, and RFID tags. The read / write industrial control computer is an embedded industrial control computer integrating a processor, memory, and storage. Biometrics are collected using multimodal identity recognition devices: facial features are captured by a visible light camera, and palm veins are collected by a near-infrared camera; electronic credentials are obtained through a QR code scanning engine or an NFC reader / writer.

[0025] Step S2 converts the collected biometric data or electronic voucher information into a unified format of identity information, synchronously linking it to environmental parameters and passenger flow density at the time of data collection. In this step, the unified format of identity information refers to converting different modalities of data into a standardized structure, including passenger identifiers, feature vectors, and timestamps; environmental parameters refer to environmental data at the time of data collection, such as light intensity, temperature, and humidity; and passenger flow density refers to the number of passengers passing through the turnstile per unit time. Scene labels are classification identifiers generated based on environmental parameters.

[0026] Data conversion is achieved through a format standardization module: biometric features are normalized and feature vectors are extracted; electronic vouchers are parsed into fields such as ticket identifiers. Environmental parameters are collected in real time by sensors, and passenger flow density is obtained by counting using infrared sensors. All data is encapsulated according to standard protocols and synchronously linked with timestamps and location information.

[0027] The association process uses the data collection time as a baseline, acquiring environmental parameters and passenger flow density before and after the data entry to form a scene triplet. Scene tags are generated according to preset rules, such as marking low light and high passenger flow as a "low light peak" scene. Data is synchronized through a dual-buffered queue, encrypted after consistency verification, and then pushed to the verification module.

[0028] This design ensures the legality of the entire data processing workflow. The original biometric image (e.g., a face RGB image of approximately 30KB / frame) resides in device memory for no more than 300 milliseconds for real-time feature extraction. Once extraction is complete (generating a 512-dimensional, approximately 2KB floating-point feature vector), the memory occupied by the original image is immediately and securely erased. All subsequent stages of the system only process, store, and compare feature vectors from these irreversible, unrecoverable original images. This technically ensures compliance with the minimum necessary principle for personal information processing.

[0029] Step S3: Perform dynamic adaptation verification on the identity information and generate scene labels based on environmental parameters. If the collected information is biometric information, extract feature vectors from the feature extraction parameter set called from the preset parameter library based on the scene label, and match them with the sub-templates of the corresponding scene in the feature template library to output the biometric verification result. If the collected information is electronic voucher information, extract the voucher identifier, determine the ticket status based on the ticket's full lifecycle data, determine spatiotemporal compliance by combining the terminal location and time period, trigger cross-verification of current and bound historical biometrics to determine the degree of consistency, and output the verification result of the fusion of ticket status, spatiotemporal compliance and degree of consistency.

[0030] Among them, dynamic adaptation verification refers to the verification mechanism that calls differentiated parameters and rules based on scene tags; preset parameter library refers to the data set that stores feature extraction parameters under different scenarios; feature template library refers to the set of biometric templates stored according to scenario classification; ticket lifecycle data refers to the full process status record of ticket from issuance to cancellation; spatiotemporal compliance refers to the legality verification of the location and time period of ticket use; cross-verification refers to the comparison and verification of electronic vouchers and bound biometric features.

[0031] The necessary procedures are as follows:

[0032] Biometric verification employs a scene-parameter mapping mechanism: scene labels serve as indexes, and corresponding feature extraction parameter sets are retrieved from a pre-defined parameter library (e.g., contrast enhancement parameters are retrieved for faces in low-light scenes, and filtering enhancement parameters are retrieved for palm veins in high-humidity scenes). After feature extraction, similarity is calculated with sub-templates of the same scene in the feature template library, and a matching threshold is dynamically determined (e.g., the threshold is reduced by 5% in peak scenes to improve the pass rate), and the verification result is output.

[0033] The electronic voucher verification employs a multi-dimensional fusion mechanism: after extracting the voucher identifier, the system queries the entire lifecycle data of the ticket to determine its status (normal / lost / expired); it combines the terminal location (entry / exit) and time period (operational / non-operational) to determine spatiotemporal compliance; and it cross-verifies the bound biometric template with the currently collected features to calculate the degree of consistency. The final verification conclusion is output by fusing these three aspects: verification is considered successful only if the ticket status is normal, spatiotemporal compliance is met, and the degree of consistency meets the standard.

[0034] Step S4: The verification results are fused. Combining scene labels, passenger flow density in the passageway, and passenger walking posture data sensed by the infrared sensor array, a comprehensive judgment value is generated through a three-level weighted algorithm of feature layer, evidence layer, and decision layer. This comprehensive judgment value is then compared with a preset threshold to generate a final judgment result. Specifically, the feature layer assigns confidence weights to the verification results based on scene labels; the evidence layer calculates the evidence support of the verification results according to the weights and preset rules; and the decision layer determines the anomaly index based on the walking posture data, determines the correction coefficient based on the passenger flow density in the passageway, and fuses the evidence support, anomaly index, and correction coefficient to generate the comprehensive judgment value. The specific process can be found in steps S41 to S44, and will not be elaborated here.

[0035] Step S5: If the overall judgment result is "pass", the gate is opened and the encrypted transaction log is recorded.

[0036] Among them, "gate opening" refers to the action of the gate controller to remove the obstruction of the passage and allow passengers to pass through; "encrypted transaction log" refers to the audit record of encrypted storage of passage transaction data (including time, station, ticket type, verification method and result, etc.).

[0037] The necessary process is as follows: The execution process adopts a dual-channel parallel mechanism: the control channel sends an opening pulse signal (lasting 200-500ms) to the gate controller via GPIO or serial port. After the gate opens, the infrared sensor monitors the passenger passage status, and automatically closes after a timeout. The data channel synchronously generates a transaction log, which includes fields such as device number, transaction serial number, ticket logical number, biometric ID (if any), transaction timestamp, and verification result. After being encrypted in SM4-CBC mode, it is stored in a local SSD dual partition (main storage and backup) and uploaded to the station computer (SC) in real time via FTP or MQ protocol. If the network is interrupted, the log is temporarily stored in a local queue and re-uploaded in batches after communication is restored to ensure the integrity and non-repudiation of transaction data.

[0038] Step S6: If the overall judgment result is "failed", a scenario-based prompt is output. In this step, the scenario-based prompt refers to a targeted prompt message generated based on the reason for failure (such as verification failure, ticket abnormality, insufficient permissions, etc.); the output methods include graphic display on the passenger screen, voice broadcast, and indicator light status switching.

[0039] The output is achieved through the UI control layer of the industrial control all-in-one machine: based on the comprehensive judgment result code of step S4 (such as biometric mismatch, expired ticket, time and space violation, etc.), the preset prompt mapping table is queried, the corresponding display resources and audio resources are called, and the passenger is provided with a clear reason for failure and guidance suggestions.

[0040] The prompting process employs a multimodal feedback mechanism: the visual channel displays a contextualized graphic interface (such as "Please try face scanning again" or "Ticket has expired, please go to the customer service center") on the passenger display screen (LCD / LED), with the interface color and icons changing according to the error type (red for prohibition, yellow for warning); the auditory channel plays corresponding voice clips through speakers (supporting TTS text-to-speech or pre-recorded audio), with the volume automatically adjusted according to ambient noise; the light indicator channel indicates the channel status through directional indicator lights (red / green) and flashing overhead lights.

[0041] In terms of technical means, the industrial control all-in-one computer's maintenance prompt rule base uses the judgment result code as an index to map to the corresponding prompt content ID, display template ID, and voice file ID. When a failure signal is received, the rule base is immediately queried to assemble a prompt data packet, which is then pushed to the passenger display screen via the HDMI / LVDS interface and driven by the audio amplifier to output the signal to the speaker. The prompt duration is dynamically set according to the scenario (generally 3-5 seconds), and the computer automatically returns to the standby interface after the timeout.

[0042] If the biometric information is facial features, feature vectors are extracted from a preset parameter library based on scene labels, and then matched with sub-templates of the corresponding scene in the feature template library. The output biometric verification results include:

[0043] Step S31: Based on the scene label, call the corresponding feature extraction parameter set from the preset parameter library.

[0044] In this step, the preset parameter library refers to the set of feature extraction parameters stored indexed by scene labels. The feature extraction parameter set refers to a combination of image processing parameters optimized for a specific scene. The method of accessing the library is to use the scene label as an index to query the preset parameter library and retrieve the optimal parameter set for the corresponding scene.

[0045] The specific calling process is as follows: The industrial control all-in-one computer uses the scene label as the query key to locate the corresponding record in the preset parameter library, extracts features to extract the parameter set (e.g., high gain parameters are called for "low light" scenes, and exposure suppression parameters are called for "strong light" scenes). The parameter library uses a hash index to achieve fast querying, and the calling results are loaded into memory for subsequent optimization.

[0046] Step S32: Optimize the called feature extraction parameter set using an environment adaptive algorithm.

[0047] In this step, the environment adaptive algorithm refers to a multimodal optimization mechanism that performs real-time correction of the feature extraction parameter set called in step S31 based on scene labels.

[0048] The algorithm is implemented by three quantization modules connected in series: environment quantizer, policy queryer, and parameter synthesizer.

[0049] Environment quantizer: Maps "scene labels" to specific physical quantity arrays, such as [light value: 150 lux, humidity: 65%].

[0050] Policy Queryer: Based on the quantized environment array, it queries a pre-defined "environment-parameter" policy table. For example, the policy table specifies that when the illumination value is in the range of 100 to 200 lux, the base brightness compensation value for face recognition is +20%, and the compensation intensity increases by 1.5 times for every 100 lux decrease in illumination.

[0051] Parameter synthesizer: Based on the base values ​​and variation rules given in the strategy table, it calculates the final set of all parameters used for feature extraction. All rules and coefficients in the strategy table are derived through regression analysis training on more than 100,000 sets of "environment-image quality optimal solution" sample data collected at different locations and time periods in subway stations, thereby enabling the algorithm to inherently adapt to the lighting and spatial characteristics of rail transit.

[0052] In addition, the following methods can be used: the industrial control all-in-one computer receives the scene label recognition result generated in step S31 and performs differentiated optimization on the called feature extraction parameter set based on the scene type: 1. Low-light scene optimization: trigger the infrared fill light array through multi-sensor linkage, match the preset color temperature correction curve (such as 3000K warm tone curve) according to the scene light intensity gradient, and simultaneously adjust the image sensor gain to the dynamic preset range (such as increasing the gain value from 1.0 to 1.5) to enhance image brightness and suppress noise. 2. Backlight scene optimization: activate the zone metering-exposure closed-loop control, apply preset short exposure parameters (such as 1 / 1000 second) to the highlight area and apply preset long exposure parameters (such as 1 / 30 second) to the shadow area, and fuse different exposure frames through the HDR synthesis algorithm to achieve dynamic contrast balance and avoid overexposure or underexposure. 3. Occlusion Scene Optimization: Key areas such as the eyes, eyebrows, and root of the nose are located using semantic segmentation algorithms. A preset dynamic weight allocation table is matched based on the occlusion area (e.g., when the occlusion area is 30%, the weight of the eye area is increased to 0.8, and the weight of other occluded areas is weakened to below 0.2). Simultaneously, feature masking technology is used to suppress interference from occluded areas, ensuring the effective proportion of computational power for feature extraction. 4. Side Face or Motion Blur Scene Optimization: A trajectory prediction-parameter pre-adjustment mechanism is activated, increasing the frame rate from the baseline value of 30fps to a scene-adaptive preset value (e.g., 60fps). Pose normalization parameters (e.g., affine transformation matrix) and motion compensation algorithms (e.g., optical flow) are used in conjunction to optimize frame quality, reducing blur caused by pose deviations or motion. After completing the above scene-specific parameter adjustments, the industrial control unit outputs the optimized feature extraction parameter set (including corrected gain, exposure, frame rate, weight allocation, and masking parameters) to the feature extraction module for feature vector extraction in step S33.

[0053] Step S33: Extract the feature vector of the face image based on the optimized feature extraction parameter set.

[0054] In this step, the optimized feature extraction parameter set refers to the combination of image processing parameters corrected by the environment adaptive algorithm in step S32, including sensor gain, exposure control, color temperature correction curve and feature weight allocation, etc.; the feature vector refers to the numerical representation extracted from the face image for identity comparison, including geometric features (facial key point coordinates, contour ratio) and texture features (local texture statistics, depth convolution features).

[0055] The necessary procedures are as follows:

[0056] Based on the optimized parameter set output in step S32, the industrial control all-in-one computer first preprocesses the face image: it applies the corrected gain and exposure parameters to adjust the image brightness, performs white balance processing through the color temperature correction curve, and masks the occluded area based on feature masking technology (such as the occluded scene determined in step S32).

[0057] Then, hierarchical feature extraction is performed:

[0058] Geometric feature extraction: MTCNN (Multi-task Convolutional Neural Network) is used to locate 5 key points on the face (center of the eyes, tip of the nose, corners of the mouth), calculate the geometric distance ratio of the facial contour (such as eye distance / face width, nose length / face length), and generate geometric feature sub-vectors; Texture feature extraction: High-dimensional texture features (usually 128-dimensional or 512-dimensional floating-point vectors) are extracted through deep convolutional networks (such as ResNet or MobileNet); or local texture statistical features are extracted using the LBP (Local Binary Pattern) algorithm.

[0059] Finally, through feature standardization (L2 normalization), the geometric feature vectors and texture feature vectors are fused into a unified feature vector according to preset weights (e.g., geometry 0.3, texture 0.7), and the output is used for template matching in step S35. The extraction process strictly follows the optimization parameters in step S32 (e.g., enabling multi-frame fusion noise reduction in low-light scenes and enabling key point weighting in occluded scenes) to ensure feature robustness and recognition accuracy.

[0060] Step S34: Retrieve the face feature sub-template corresponding to the scene label from the feature template library.

[0061] In this step, the feature template library refers to the collection of registered face feature templates stored according to scene labels, including baseline templates under different lighting, angles, and occlusion conditions; the face feature sub-template refers to a specific scene template subset that matches the current scene label; the calling method refers to the process of accurately retrieving the corresponding template subset from the feature template library using the scene label as the index key.

[0062] The call is implemented through a scene-template mapping mechanism: using the scene labels generated in step S31 (such as "low light front" and "strong light side") as query conditions, the pre-stored feature template data for that scene is located from the feature template library and loaded.

[0063] The calling process is as follows: The industrial control all-in-one machine uses the scene label as the index key to query the scene-template mapping table in the feature template library. The feature template library uses structured storage (such as a database or in-memory hash table) and organizes template data by scene label (e.g., "low light" scene stores enhanced templates registered under low light conditions, "occlusion" scene stores templates for key eye areas). The system accurately retrieves the matching facial feature sub-templates (containing the feature vector set and metadata of registered users in that scene) based on the scene label and loads them into the memory buffer. For example, if the scene label is "low light peak", the system calls the low light optimized template subset corresponding to this label, excluding strong light or normal light templates, ensuring that the matching calculation in subsequent step S35 is performed in similar scene templates, improving matching accuracy and speed.

[0064] Step S35: The extracted feature vector is matched with the called sub-template to calculate the similarity value. In this step, the feature vector refers to the numerical representation of the facial geometric and texture features extracted in step S33; the sub-template refers to the registered facial feature template corresponding to the current scene label called from the feature template library in step S34; the similarity value refers to the quantitative index (range 0-1, the higher the value, the higher the matching degree) that measures the degree of matching between the two.

[0065] The matching method employs multi-dimensional similarity calculation: by using algorithms such as cosine similarity or Euclidean distance, the spatial distance or cosine value of the angle between the extracted feature vector and the feature vector of the sub-template is calculated and mapped to a similarity value.

[0066] The matching process is as follows:

[0067] The feature vector output in step S33 is compared with the sub-template called in step S34. For each set of registered feature vectors in the sub-template, the system calculates its similarity to the currently extracted feature vector:

[0068] The cosine similarity formula can be used: calculate the cosine of the angle between two vectors, and the closer the value is to 1, the more consistent the directions are; or the Euclidean distance formula can be used: calculate the distance in the vector space, and map it to a similarity value after normalization.

[0069] The system iterates through all template vectors in the sub-templates, taking the highest similarity value as the final similarity output (or taking the Top-K mean), which is then compared with the dynamic threshold in step S36 to determine whether the match passes or fails. The calculation process strictly follows the optimized feature weight allocation in step S32 (such as increasing the weight of eye features in occluded scenes) to ensure scene adaptability.

[0070] Step S36: Dynamically determine the matching threshold based on the scene label, compare the similarity value with the matching threshold, and output the face feature verification result. In this step, the matching threshold refers to the similarity threshold used to determine whether a face feature passes verification; dynamic determination refers to the mechanism of adjusting the threshold size in real time based on the scene label (e.g., lowering the threshold in peak scenes to improve the pass rate, and raising the threshold in strong light scenes to reduce the false recognition rate); the verification result refers to the pass / fail judgment and confidence level generated based on the comparison.

[0071] The necessary procedures are as follows:

[0072] First, the preset threshold adjustment rule table is queried based on the scene label. The rule table stores the baseline threshold and adjustment coefficient according to the scene type: for example, the baseline threshold of 0.80 is used for the "normal" scene, it is lowered by 5% to 0.76 for the "low light peak" scene, it is raised to 0.85 for the "strong light sparse" scene, and the "occlusion" scene is dynamically fine-tuned according to the occlusion area.

[0073] The system then compares the similarity value output in step S35 with the adjusted matching threshold: if the similarity value is greater than or equal to the threshold, the face feature matching is deemed successful, and the verification result and similarity confidence are output; if the similarity value is less than the threshold, the matching is deemed unsuccessful, and the verification failure result and failure reason code (such as "insufficient similarity") are output.

[0074] The final verification results (including pass / fail flags, similarity values, scene labels, and threshold information) are packaged into a standard format for comprehensive judgment and fusion in step S4. This dynamic threshold mechanism ensures a reasonable balance between pass rate and security under different environmental conditions.

[0075] If the biometric information is palm vein features, the feature vector is extracted from the feature extraction parameter set called from the preset parameter library based on the scene label, and matched with the corresponding scene sub-template in the feature template library. The output biometric verification results include:

[0076] Step S3A: Based on the scene label, call the corresponding feature extraction parameter set from the preset parameter library.

[0077] In this step, scene labels refer to classification labels (such as "high humidity", "uneven light source", "angle deviation") generated based on palm vein acquisition environment parameters (including light source uniformity, palm humidity, type of obstruction, shooting angle and imaging distance); preset parameter library refers to the set of palm vein feature extraction parameters stored according to scene label index, including parameters such as vein texture enhancement, filtering and noise reduction, ROI boundary setting; feature extraction parameter set refers to the combination of image processing parameters optimized for specific scenes.

[0078] The calling process is as follows: The scene labels generated in step S2 (including environmental parameters such as light source uniformity, humidity, and occlusion type) are used as the query key to locate the corresponding record in the preset parameter library and extract the palm vein feature extraction parameter set. For example, the "high humidity" scene calls the deblurring enhancement and contrast enhancement parameters; the "uneven light source" scene calls the multi-region brightness equalization parameters; the "angle deviation" scene calls the affine transformation correction parameters; and the "distance anomaly" scene calls the scaling compensation parameters. The parameter library uses a hash index for fast lookup, and the calling results are loaded into memory for palm vein image preprocessing and feature extraction in step S3B.

[0079] Step S3B: Extract the feature vector of the palm vein image based on the called feature extraction parameter set.

[0080] The feature extraction parameter set called refers to the image processing parameters obtained in step S3A based on scene labels (such as high humidity, uneven light source, and angle deviation), including multi-region brightness equalization intensity, deblurring filter coefficient, vein texture enhancement weight, and ROI (region of interest) boundary setting; the palm vein image refers to the palm vein image acquired by a near-infrared camera in step S1 and preprocessed according to the above parameter set; the feature vector refers to the numerical representation extracted from the palm vein image, including vein texture features (main trunk, branches, and minutiae point structure) and palm geometric features (contour, finger root key points, and size ratio).

[0081] Based on the feature extraction parameter set called in step S3A, perform layered extraction on the palm vein image:

[0082] 1. Preprocessing Enhancement: Targeted parameter optimization is applied based on scene labels. For example, in a "high humidity" scene, a deblurring enhancement strategy is triggered, contrast enhancement parameters are applied, and a vein texture enhancement algorithm is used to highlight the vascular structure; in a "uneven light source" scene, a multi-region brightness equalization algorithm is used to correct the brightness deviation of the palm vein area according to a preset illumination gradient; in a "angle deviation" scene, an affine transformation correction model is called to project the tilted palm onto a standard coordinate system and adjust the ROI boundary. 2. Vein Texture Feature Extraction: Hessian matrix multi-scale filtering or Gabor filtering algorithms are used to extract the texture structure information of the main trunk, branches, and minutiae points (endpoints, bifurcation points) of the palm vein vascular network, generating texture feature sub-vectors. 3. Geometric Feature Extraction: Through palm contour detection and key point localization (palm center, finger roots), the geometric dimensions and shape parameters of the palm are calculated, generating geometric feature sub-vectors. 4. Feature Fusion and Standardization: The vein texture feature vector and the geometric feature vector are fused according to preset weights (e.g., texture 0.7, geometry 0.3). After L2 normalization, a feature vector with a unified format is output for template matching in step S3D. This process strictly follows the scene-specific parameters of step S3A (e.g., increasing the weight of the effective region in occluded scenes) to ensure feature robustness.

[0083] Step S3C: Retrieve the palm vein feature sub-template corresponding to the scene label from the feature template library.

[0084] The feature template library refers to the collection of registered palm vein feature templates categorized and stored according to scene labels, covering baseline templates under different humidity, lighting, angle, and distance conditions; the palm vein feature sub-templates refer to a specific subset of scene templates that match the current scene label (such as "high humidity", "uneven light source", "angle deviation"), including vein texture feature templates and palm geometric feature templates; the calling method refers to the process of accurately retrieving the corresponding template subset from the feature template library using the scene label as the index key.

[0085] The calling process is as follows:

[0086] The industrial control all-in-one computer uses scene tags as index keys to query the scene-template mapping table in the feature template library. The feature template library uses structured storage, organizing template data according to scene tags (e.g., the "high humidity" scene stores templates that have undergone deblurring and enhancement processing, the "uneven light source" scene stores templates after brightness equalization, and the "angle deviation" scene stores templates after affine transformation correction).

[0087] The system accurately retrieves the matching palm vein feature sub-template (containing the vein texture feature vector, geometric feature vector, and metadata of the registered user in that scene) based on the scene label and loads it into the memory buffer. For example, if the scene label is "high humidity", the system calls the humidity-optimized template subset corresponding to that label, excluding normal humidity templates, to ensure that the subsequent S3D matching calculations are performed on similar scene templates, thereby improving matching accuracy and speed.

[0088] In step S3D, the extracted feature vectors are matched with the called sub-templates to calculate the similarity value.

[0089] In this step, the feature vector refers to the fusion representation of the palm vein texture features extracted in step S3B and the palm geometric features; the sub-template refers to the registered palm vein feature template called in step S3C that corresponds to the current scene label (such as high humidity, uneven light source); and the similarity value refers to a comprehensive quantitative index (range 0-1) that measures the degree of matching between the two.

[0090] The matching process is as follows:

[0091] The industrial control integrated computer performs a hierarchical comparison between the feature vector output in step S3B and the sub-template called in step S3C:

[0092] Texture feature matching: The similarity value between the extracted vein texture features (main vein and branch structure) and the texture vector of the sub-template is calculated using cosine similarity or correlation coefficient algorithms.

[0093] Geometric feature matching: The similarity value between the geometric features of the palm (contour, finger root key points) and the geometric vector of the sub-template is calculated using Euclidean distance or Hamming distance algorithm;

[0094] Weighted fusion: Based on preset weight allocation rules (such as texture feature weight 0.7, geometric feature weight 0.3, or dynamically adjusted according to scene labels), the texture similarity and geometric similarity are weighted and fused to generate the final similarity value.

[0095] The system iterates through all template vectors in the sub-templates, taking the highest similarity value (or Top-K mean) as the output, which is then compared with the dynamic threshold in step S3E to determine whether the match passes or fails. The calculation process follows the scene-specific parameters of step S3A (such as increasing the weight of the effective region in occluded scenes) to ensure scene adaptability.

[0096] In step S3E, the matching threshold is dynamically determined based on the scene label, the similarity value is compared with the matching threshold, and the palm vein feature verification result is output.

[0097] The matching threshold refers to the similarity threshold used to determine whether the palm vein feature passes the verification; dynamic determination refers to the mechanism of adjusting the threshold size in real time according to scene labels (such as "high humidity", "uneven light source", "angle deviation", "distance anomaly"). For example, the threshold is lowered in high humidity scenes to compensate for the decrease in image quality, and the fault tolerance range is adjusted in distance anomaly scenes; the verification result refers to the pass / fail judgment and confidence level generated based on the comparison.

[0098] The comparison process is as follows:

[0099] The industrial control all-in-one computer first queries the preset threshold adjustment rule table based on scene tags (including light source uniformity, humidity, occlusion type, etc.). The rule table stores adjustment strategies according to palm vein-specific scene types: for example, the threshold may be lowered by 5%-10% in the "high humidity" scene to improve the pass rate; the "uneven light source" scene is dynamically fine-tuned based on the degree of brightness deviation; the "angle deviation" scene sets the attitude tolerance threshold; and the "distance anomaly" scene adjusts the fault tolerance range based on the accuracy loss after scaling compensation.

[0100] The system then compares the similarity value output in step S3D with the adjusted matching threshold: if the similarity value is greater than or equal to the threshold, the palm vein feature matching is deemed successful, and the verification result and similarity confidence are output; if the similarity value is less than the threshold, the matching is deemed unsuccessful, and the verification failure result and failure reason code (such as "vein feature mismatch") are output.

[0101] The final verification results (including pass / fail flags, similarity values, scene labels, and threshold information) are packaged into a standard format for comprehensive judgment and fusion in step S4. This dynamic threshold mechanism ensures a reasonable balance between pass rate and safety under different humidity, lighting, angle, and distance conditions.

[0102] If the biometric information includes palm vein features and facial features, feature vectors are extracted from the feature extraction parameter set called from the preset parameter library based on scene labels, and matched with the corresponding scene sub-templates in the feature template library. The output biometric verification results include: Step S3a, generating joint scene labels based on environmental parameters. The environmental parameters include a first type of parameter related to facial features and a second type of parameter related to palm vein features. In this step, the joint scene label refers to a comprehensive classification identifier that simultaneously covers the facial and palm vein acquisition environment, used to characterize the composite environmental conditions during dual-modal feature extraction. The first type of parameter refers to environmental parameters that affect the quality of facial images, including light intensity, occlusion area, facial angle, and time information. The second type of parameter refers to environmental parameters that affect the quality of palm vein images, including light source uniformity, palm humidity, type of occlusion, shooting angle, and imaging distance.

[0103] The necessary procedures are as follows:

[0104] The industrial control all-in-one machine obtains environmental parameters from step S2 and divides them into two categories: the first category of parameters (face-related: light intensity, occlusion area, angle, time period) and the second category of parameters (palm vein-related: light source uniformity, humidity, occlusion type, shooting angle, imaging distance).

[0105] Based on preset joint scene classification rules (such as decision trees or rule mapping tables), the system combines and judges the two types of parameters. For example, if the first type of parameter indicates "low light" and the second type of parameter indicates "high humidity", then a joint scene label of "low light and high humidity" is generated; if the first type is "strong light" and the second type is "angle deviation", then a joint label of "strong light and angle deviation" is generated.

[0106] The generated joint scene label simultaneously indexes the parameter optimization strategies for both face and palm vein, serving as a unified index for step S3b to call, ensuring that dual-modal feature extraction is performed under a collaboratively optimized environment configuration.

[0107] Step S3b: Based on the joint scene label, call the face feature extraction parameter set and the palm vein feature extraction parameter set from the preset parameter library respectively;

[0108] In this step, the joint scene label refers to the composite identifier generated in step S3a (such as "low light and high humidity" and "strong light angle deviation"), which covers both the face and palm vein acquisition environment; the preset parameter library refers to the set of dual-modal feature extraction parameters stored according to the joint scene label index, which includes the optimization parameter branches for each of the face and palm vein; the face feature extraction parameter set refers to the face image processing parameters (including exposure control, gain adjustment, HDR synthesis parameters, etc.) for lighting and occlusion conditions in the joint scene; the palm vein feature extraction parameter set refers to the palm vein image processing parameters (including brightness equalization, deblurring filtering, ROI correction parameters, etc.) for humidity and light source conditions in the joint scene.

[0109] The calling process is as follows:

[0110] The industrial control all-in-one computer uses the joint scene tag (such as "low light and high humidity") generated in step S3a as the query key to perform a dual-branch search in the preset parameter library:

[0111] Face parameter retrieval: Based on the first type of parameters (light intensity, occlusion area) in the joint label, locate the face parameter branch and call the corresponding feature extraction parameter set (such as infrared supplementary light parameters, high gain settings, and color temperature correction curves in low light scenes).

[0112] Palm vein parameter retrieval: Based on the second type of parameters (humidity, light source uniformity) in the joint label, locate the palm vein parameter branch and call the corresponding feature extraction parameter set (such as contrast enhancement parameters and vein texture enhancement filter coefficients in high humidity scenes).

[0113] The system loads the two sets of parameters retrieved into memory buffers, marking them as the initial parameters for the face processing channel and the palm vein processing channel, respectively, for optimization by step S3c using the environment adaptation algorithm. This separate invocation mechanism ensures that the two modalities start under their respective optimal initialization conditions, while maintaining consistency in environmental perception through joint scene labels.

[0114] Step S3c involves optimizing the parameter sets for face feature extraction and palm vein feature extraction using an environment adaptive algorithm. The specific process is detailed in steps S3c1 to S3c4 and will not be elaborated upon here.

[0115] Step S3d: Extract the feature vector of the face image based on the optimized face feature extraction parameter set, extract the feature vector of the palm vein image based on the optimized palm vein feature extraction parameter set, and fuse the feature vector of the face image and the feature vector of the palm vein image to obtain a joint feature vector.

[0116] Parallel execution of dual-modal feature extraction by the industrial control all-in-one computer:

[0117] Facial Feature Extraction: Based on the optimized facial parameter set (such as low-light compensation and occlusion mask weights), the facial images captured by the visible light camera are preprocessed. MTCNN is used to locate facial key points and extract geometric features, and a deep convolutional network (or LBP) is used to extract texture features. After L2 normalization, a facial feature vector is formed.

[0118] Palm vein feature extraction: Based on the optimized palm vein parameter set (such as high-humidity deblurring and light source equalization), the palm vein images acquired by the near-infrared camera are preprocessed. Hessian matrix multi-scale filtering is used to extract vein texture features (main trunk, branches, and minutiae points), which are combined with palm geometric contour features and normalized to form a palm vein feature vector.

[0119] Feature fusion: The facial feature vector and the palm vein feature vector are fused according to a preset fusion rule (such as direct concatenation or weighted superposition, with weights dynamically adjusted based on the joint scene label, such as increasing the weight of palm veins in high humidity scenes and increasing the weight of faces in low light scenes) to generate a joint feature vector. This vector contains biometric information of both the face and palm veins, which is used for multimodal template matching in the subsequent step S3e.

[0120] Step S3e: Retrieve the face feature sub-template and palm vein feature sub-template corresponding to the joint scene label from the feature template library.

[0121] In this step, the feature template library refers to the set of bimodal registered feature templates stored according to the joint scene label, including the face feature template branch and the palm vein feature template branch; the face feature sub-template refers to the subset of face feature templates that match the first type of parameters (lighting, occlusion, etc.) in the joint scene label; the palm vein feature sub-template refers to the subset of palm vein feature templates that match the second type of parameters (humidity, light source, etc.) in the joint scene label.

[0122] The calling method adopts a joint index dual-branch retrieval: using the joint scene label generated in step S3a as the global index key, the face template branch and palm vein template branch are queried in parallel in the feature template library, and the dual-modal feature sub-templates that match the current composite environment conditions are extracted respectively.

[0123] The calling process is as follows:

[0124] The industrial control all-in-one computer will use the combined scene tag (such as "low light and high humidity") as the query key to perform a two-branch search in the feature template library:

[0125] Face sub-template retrieval: Based on the first type of parameters (light intensity, occlusion area, angle) in the joint label, locate the face template branch and call the face feature sub-template of the corresponding scene (such as the enhanced face template registered in low light scene, and the eye focus template in occlusion scene).

[0126] Palm vein sub-template retrieval: Based on the second type of parameters (humidity, light source uniformity, angle) in the joint label, locate the palm vein template branch and call the palm vein feature sub-template of the corresponding scene (such as the deblurred vein template in the high humidity scene, and the brightness equalization template in the uneven light source scene).

[0127] The system loads the retrieved facial feature sub-templates and palm vein feature sub-templates into memory buffers, marking them as the comparison benchmarks for the dual-modal matching channels, for use in step S3f to perform multi-dimensional similarity calculation of the joint feature vectors. This joint invocation mechanism ensures that dual-modal comparison is performed in template libraries adapted to their respective scenarios, improving the joint recognition accuracy in complex environments.

[0128] Step S3f involves matching the joint feature vector with the invoked face feature sub-template and palm vein feature sub-template to obtain a comprehensive similarity value. The joint feature vector refers to the unified feature representation output in step S3d that integrates face and palm vein information (which can be divided into face feature components and palm vein feature components); the face feature sub-template and palm vein feature sub-template refer to the set of registered templates corresponding to the two types of biometric features, invoked in step S3e based on the joint scene label; the comprehensive similarity value refers to the final quantitative index (range 0-1) after fusing the dual-modal matching results.

[0129] The matching method employs multi-dimensional separate calculation and weighted fusion: the joint feature vector is independently calculated with the face sub-template and the palm vein sub-template, and then the similarity of the two dimensions is weighted and fused into a comprehensive similarity value according to the weight allocation rules determined by the joint scene label.

[0130] The industrial control integrated computer decomposes the joint feature vector output in step S3d into facial feature components and palm vein feature components, and compares them with the corresponding sub-templates called in step S3e:

[0131] Face dimension matching: Calculate the similarity between the face feature components and each template in the face feature sub-template (using cosine similarity or Euclidean distance algorithm), and take the highest similarity value or Top-K mean as the face similarity value. Palm vein dimension matching: Calculate the similarity between the palm vein feature components and each template in the palm vein feature sub-template to obtain the palm vein similarity value. Weighted fusion: Based on the joint scene label query and the preset weight allocation rules (e.g., palm vein weight is increased in "low light and high humidity" scene to compensate for the decrease in face imaging quality, and face weight is higher in "normal" scene), obtain the face weight coefficient and palm vein weight coefficient, and sum the two similarity values ​​with weights to generate a comprehensive similarity value (e.g., comprehensive similarity = 0.4 × face similarity + 0.6 × palm vein similarity).

[0132] This comprehensive similarity value is used for comparison with the dynamic threshold in step S3g, and the final biometric verification result is output.

[0133] Step S3g: Dynamically determine the matching threshold based on the joint scene label, compare the comprehensive similarity value with the matching threshold, and output the biometric verification result.

[0134] The comparison process is as follows: The industrial control all-in-one machine queries the preset threshold adjustment rule table based on the joint scene label. The rule table stores differentiated threshold strategies for dual-modal composite scenes: for example, in the "low light and high humidity" scene, the threshold may be lowered to compensate for insufficient face imaging while ensuring the reliability of the palm vein weight increase; in the "strong light angle deviation" scene, the fault tolerance range is adjusted according to the dual-modal confidence balance.

[0135] The system compares the overall similarity value output in step S3f with the adjusted matching threshold: if the overall similarity value is greater than or equal to the threshold, the bimodal biometric matching is deemed successful, and the verification result and overall confidence level are output; if the overall similarity value is less than the threshold, the matching is deemed unsuccessful, and the verification failure result and failure reason code (such as "bimodal feature mismatch") are output.

[0136] The final verification result (including pass / fail flags, comprehensive similarity value, joint scene label, and threshold information) is packaged into a standard format for comprehensive judgment and fusion in step S4. This joint dynamic threshold mechanism ensures that bimodal verification can maintain a reasonable balance between security and pass rate even under complex and composite environmental conditions.

[0137] The facial feature extraction parameter set and the palm vein feature extraction parameter set were optimized using an environment adaptive algorithm, including:

[0138] Step S3c1: Obtain the current operating mode of the integrated industrial control unit (ICM) and the environmental parameters in the combined scenario tag. In this step, the current operating mode refers to the real-time operating status of the ICM, which is read from the device status register (normal service / maintenance / degradation mode).

[0139] The industrial control all-in-one computer performs the following parallel operations:

[0140] Operating mode acquisition: Access the device status register (such as the system control register or mode flag), read the current operating mode identifier (such as 0x01 for normal mode, 0x02 for maintenance mode, and 0x04 for degraded mode), and synchronously record the mode switching timestamp as the algorithm input parameter.

[0141] Environmental parameter acquisition: Analyze the joint scene label data structure generated in step S3a, extract the first type of parameters (face-related: light intensity value, occlusion area ratio, face deflection angle, time period information) and the second type of parameters (palm vein-related: light source uniformity index, palm humidity value, occlusion type code, shooting angle deviation, imaging distance value), and combine them into an environmental parameter matrix.

[0142] The acquired current working mode and environmental parameter matrix are used as the first and second input parameters of the environment adaptive algorithm, and are synchronously input into step S3c2 for subsequent benchmark parameter set extraction and dynamic parameter generation.

[0143] Step S3c2: Based on the current working mode, obtain the face reference parameter set and palm vein reference parameter set from the preset mode-parameter mapping relationship respectively.

[0144] Among them, the mode-parameter mapping relationship refers to the preset parameter configuration table indexed by working mode, which is stored in the local database of the industrial control computer; the face reference parameter set and the palm vein reference parameter set refer to the default initial parameters for feature extraction in each working mode (including exposure time, gain, fill light intensity, ROI boundary, etc.).

[0145] The necessary procedures are as follows:

[0146] The industrial control all-in-one computer queries the mode-parameter mapping table based on the current working mode identifier read in step S3c1:

[0147] If the current service mode is normal, extract the default performance parameters: the face baseline parameter set includes an exposure time of 10ms, a gain of 1.0, a frame rate of 30fps, and standard HDR off; the palm vein baseline parameter set includes an infrared fill light intensity of 80%, ROI standard boundary, and default filter coefficients.

[0148] If the current mode is maintenance, extract the desensitization parameters: adjust the face baseline parameters to exposure 5ms and gain 0.8 to reduce false triggering; adjust the palm vein baseline parameters to infrared intensity 60% and reduce the ROI range for faster detection.

[0149] If the current mode is degraded (e.g., due to equipment aging or partial sensor failure), extract the fault tolerance parameters: enable the backup algorithm path, reduce the resolution requirement, and enable the multi-frame fusion backup mechanism.

[0150] The mapping table uses structured storage (such as an SQLite database or JSON configuration file), with working mode encoding as the primary key, and establishes a hash index with O(1) time complexity to ensure that the baseline parameter set is quickly loaded into the memory buffer for subsequent steps of S3c3 and dynamic environmental parameters.

[0151] Step S3c3: Based on environmental parameters, generate a dynamic parameter set for the face and a dynamic parameter set for the palm vein using a preset environmental feature adaptation function.

[0152] In this step, the environmental parameters refer to the specific data matrix parsed from the joint scene labels in step S3c1, including the first type of parameters related to the face (light intensity, occlusion area, angle, time period) and the second type of parameters related to the palm vein (light source uniformity, humidity, occlusion type, shooting angle, imaging distance); the environmental feature adaptation function refers to the preset nonlinear calculation rule (such as piecewise linear function or lookup table method) that maps the environmental parameters to compensation parameters; the face dynamic parameter set and the palm vein dynamic parameter set refer to the compensation adjustment parameters (including exposure compensation value, gain correction coefficient, filter intensity, etc.) calculated according to real-time environmental disturbances.

[0153] The necessary procedures are as follows:

[0154] The industrial control integrated computer inputs the environmental parameter matrix obtained in step S3c1 into the preset environmental feature adaptation function:

[0155] Face dynamic parameter generation: Compensation values ​​are calculated based on the first type of parameters. For example, when the illumination intensity L < 50 lux, the exposure compensation value ΔE = 1.5 × (50 - L) / 50 is calculated using the adaptation function, and the gain dynamic coefficient G = 1.2; when the occlusion area A > 20%, the weight enhancement value W for the eye region is calculated as 0.8, and the occlusion area mask threshold M = 0.2; when the angle deviation θ > 15°, the pose normalized rotation compensation matrix R is calculated. The output face dynamic parameter set {ΔE, G, W, M, R} is then generated.

[0156] Palm vein dynamic parameter generation: Compensation values ​​are calculated based on the second type of parameters. For example, when humidity H > 70%, the deblurring filter coefficient F = 0.8 × H / 100 and the contrast enhancement value C = 1.3 are calculated through the adaptation function; when the light source uniformity U < 0.6, the multi-region brightness equalization intensity B = 1.2 × (1 - U) is calculated; when the imaging distance D ≠ standard value, the scaling compensation coefficient S = D / standard distance is calculated. The output palm vein dynamic parameter set {F, C, B, S} is generated.

[0157] After the two dynamic parameter sets are generated synchronously, they are input together with the baseline parameter set obtained in step S3c2 into step S3c4 for weighted fusion optimization.

[0158] Step S3c4: According to the preset priority rules, the face baseline parameter set and the face dynamic parameter set are weighted and fused to obtain the optimized face feature extraction parameter set, and the palm vein baseline parameter set and the palm vein dynamic parameter set are weighted and fused to obtain the optimized palm vein feature extraction parameter set.

[0159] In this step, the preset priority rule refers to the strategy for allocating the weights of the baseline parameter and the dynamic parameter (e.g., baseline parameter weight α=0.6, dynamic parameter weight β=0.4); weighted fusion refers to the calculation process of linearly superimposing the two types of parameters according to the weights.

[0160] The fusion method adopts dual-channel independent weighting: weighting operations are performed separately for the face and palm vein parameters according to preset priority rules to generate an optimized feature extraction parameter set.

[0161] The necessary procedures are as follows:

[0162] The industrial control all-in-one computer first queries the preset priority rules to obtain the baseline parameter weight coefficient α and the dynamic parameter weight coefficient β (usually α+β=1).

[0163] Then, a dual-path parallel fusion is performed: 1. Face parameter fusion: According to the formula "Optimized face parameters = baseline weight × face baseline parameter set + dynamic weight × face dynamic parameter set", the face baseline parameter set from step S3c2 and the face dynamic parameter set from step S3c3 are weighted and fused; 2. Palm vein parameter fusion: Similarly, according to the formula "Optimized palm vein parameters = baseline weight × palm vein baseline parameter set + dynamic weight × palm vein dynamic parameter set", the palm vein baseline parameter set from step S3c2 and the palm vein dynamic parameter set from step S3c3 are weighted and fused.

[0164] After fusion, the system performs a preset range check and corrects parameter values ​​that exceed the reasonable range (such as limiting the upper and lower thresholds of gain). It then outputs the optimized face feature extraction parameter set and the optimized palm vein feature extraction parameter set for step S3d to extract dual-modal feature vectors.

[0165] The feature layer assigns confidence weights to the verification results based on scene labels; the evidence layer calculates the evidence support of the verification results according to the weights and preset rules; the decision layer determines the anomaly index based on traffic posture data, determines the correction coefficient based on the passenger flow density of the passage, and integrates the evidence support, anomaly index, and correction coefficient to generate a comprehensive judgment value, including:

[0166] In step S41, the feature layer queries the preset feature weight configuration table based on the scene labels to assign corresponding confidence weights to the verification results. Here, the feature layer refers to the first layer of the three-level weighting algorithm, responsible for initially assigning confidence weights to the verification results; the scene labels refer to the classification identifiers generated in step S2 based on environmental parameters (lighting, passenger flow density, etc.) (e.g., "low light peak," "strong light sparse"); the feature weight configuration table refers to the preset weight allocation rule base indexed by scene labels; the confidence weight refers to the reliability coefficient (range 0-1) assigned to the verification results, reflecting the credibility of the verification results in the current scene; and the verification results refer to the biometric verification results (face / palm vein / combined) or electronic voucher verification results output in step S3.

[0167] The fundamental innovation of the "feature-evidence-decision" three-layer model designed in this invention lies in the deep parameterization binding of the general decision-making architecture with the domain knowledge of rail transit gate passage.

[0168] The task of the feature layer is to quantify the scene's impact into initial weights. For example, in system configuration, a "strong backlight" scene might cause the initial weight of the face verification result to be reduced by 30%.

[0169] The core of the evidence layer is a built-in set of quantitative rules to resolve conflicts between biometric features and electronic credential verification results.

[0170] The key to the decision-making layer is the introduction of operational indicators such as passenger flow density (people / minute) to dynamically adjust the release threshold. All the parameters, including weights, rules, and thresholds across all layers, totaling thousands, are not set manually based on experience. Instead, they are trained using a multi-objective optimization algorithm on a dataset containing over one million real subway gate passage records (covering all scenarios such as normal passage, tailgating, and ticketing anomalies). Therefore, this model is essentially a dedicated decision engine driven by rail transit operational data.

[0171] The necessary procedures are as follows:

[0172] The industrial control all-in-one computer queries the preset feature weight configuration table based on the scene tags generated in step S2. The configuration table stores the weight allocation strategy according to scene type:

[0173] In normal scenarios (sufficient lighting, moderate passenger flow), the confidence weights for both biometric verification and electronic voucher verification are set to 1.0 (or 0.5 each, summed to 1.0). In low-light peak scenarios (insufficient lighting, dense passenger flow), the weight of biometric verification is reduced (e.g., reduced to 0.4), while the weight of electronic voucher verification is increased (e.g., increased to 0.6) to compensate for the uncertainty in biometrics caused by decreased image quality. In strong, sparse-light scenarios, the weight of biometric verification is increased (e.g., 0.6), while the weight of electronic voucher verification is reduced (e.g., 0.4), prioritizing biometric recognition.

[0174] For bimodal biometrics (face + palm vein), the weights are further subdivided: for example, in the "low light and high humidity" scenario, the weight of the face is reduced to 0.3, while the weight of the palm vein is increased to 0.7, so as to make full use of the characteristic that the palm vein is not sensitive to light.

[0175] The system assigns the confidence weights obtained from the query to the corresponding verification results, forming a mapping pair of (verification result type, confidence weight), and outputs it to step S42 for the evidence layer to calculate the evidence support.

[0176] In step S42, the evidence layer processes the verification results according to the assigned confidence weights and calls the preset evidence fusion and conflict resolution rule base for comprehensive calculation to generate a unified evidence support level.

[0177] In this step, the evidence layer refers to the middle layer of the three-level weighting algorithm, which is responsible for combining the confidence weights assigned by the feature layer with the verification results and generating a unified evidence support degree through logical operations; the verification result refers to the biometric verification conclusion (including similarity value) or electronic voucher verification conclusion (including ticket status, spatiotemporal compliance, and consistency) output in step S3; the confidence weight refers to the weight coefficient assigned by step S41 according to the scene label; the evidence fusion and conflict resolution rule base refers to the preset rule set, including DS evidence synthesis rules, Bayesian fusion rules, and consistency calibration rules, which are used to handle the fusion of multi-source verification results or the confidence calibration of a single verification result; the unified evidence support degree refers to the scalar value (range 0-1) output after calculation by the rule base, which characterizes the overall credibility of the verification results.

[0178] The processing procedure is as follows:

[0179] The industrial control integrated computer receives the confidence weight assigned in step S41 and the verification result in step S3.

[0180] First, identify the verification type of the verification result (biometric verification or electronic certificate verification), analyze its generation process, and construct a multi-dimensional credibility vector that reflects key credibility factors (such as biometric similarity credibility, ticket status credibility, spatiotemporal compliance credibility, etc.).

[0181] Next, the confidence weights assigned to the feature layer are applied to the confidence vector as weighting coefficients.

[0182] Subsequently, a pre-defined evidence fusion and conflict resolution rule base is invoked: for a single verification result, the deviation from the scenario benchmark credibility is calculated and corrected through the consistency calibration module, and the comprehensive credibility is output; for a dual verification result of biometrics and electronic credentials, the joint support is calculated using the DS evidence synthesis rule, and the evidence conflict coefficient is calculated simultaneously. If a conflict exists, it is resolved using conflict resolution rules (such as weighted reconciliation or priority adjudication). Finally, a Bayesian posterior fusion or normalization algorithm is used to nonlinearly synthesize the processed credibility and confidence weights to generate a unified evidence support (scalar value, range 0-1), which is output to step S43 for use by the decision-making layer.

[0183] In step S43, the decision-making layer calculates three quantitative indicators based on the passage posture data perceived by the infrared sensor array: abnormal dwell time, deviation of movement trajectory, and conflict degree of behavior pattern. These indicators are then aggregated through a preset posture risk assessment model to generate an anomaly index.

[0184] Among them, the abnormal dwell time refers to the deviation of a passenger's dwell time in the passage exceeding the normal passage threshold; the deviation of the movement trajectory refers to the degree of deviation of the passenger's actual movement path from the standard passage path (such as the center line of the passage); the degree of behavioral pattern conflict refers to the degree of difference between the passenger's behavior (such as sudden turning back, multiple people walking side by side) and the normal single passage mode; the posture risk assessment model refers to the preset algorithm model (such as weighted summation or fuzzy logic model) used to aggregate the above three indicators into a single risk value; and the anomaly index refers to the quantified passenger passage risk value (range 0-1, the higher the value, the more abnormal the risk).

[0185] The industrial control all-in-one computer receives real-time passage posture data (including timestamps of passengers entering / leaving the passage, real-time location coordinates, and moving speed) collected by the infrared sensor array.

[0186] Based on this data, the system calculates three quantitative indicators: 1. Dwell time anomalies: Calculates the total time a passenger spends from entering to leaving the passage (or the dwell time in a specific area), compares it with a preset normal passage time threshold (e.g., 5 seconds), and calculates the degree of anomaly for the excess portion (e.g., exceeding 15 seconds is marked as high anomaly); 2. Movement trajectory deviation: Draws the actual movement trajectory based on the position coordinate sequence, compares it with the standard passage trajectory (passage centerline), and calculates the lateral deviation distance or trajectory tortuosity (e.g., calculates the root mean square error between the actual path and the straight path); 3. Behavioral pattern conflict degree: Analyzes passenger behavior sequences (e.g., detects sudden stops, reverse movements, excessive spacing between multiple people, etc.), matches them with a preset normal behavior pattern (single person passing at a constant speed), and quantifies the probability of conflict or the frequency of abnormal behavior. The above three quantitative indicators are input into a preset posture risk assessment model, and weighted and aggregated according to preset weights (e.g., 40% for dwell time anomalies, 30% for movement trajectory deviation, and 30% for behavioral pattern conflict degree) to generate an anomaly index (range 0-1). This index represents the degree of abnormal risk of passenger passage behavior (e.g., an index of 0.8 indicates a high probability of gate breaking or tailgating), and is output to step S44 for the decision-making layer to integrate and use.

[0187] In step S44, the decision-making layer also calculates the correction coefficient based on the passenger flow density of the passage, combined with the current time period attributes and passage type, through a preset operational pressure quantification model; the decision-making layer then inputs the evidence support, anomaly index and correction coefficient into a preset multi-source risk decision function; the function performs fusion calculation on the input evidence support, anomaly index and correction coefficient according to a preset fusion strategy, and outputs a comprehensive judgment value.

[0188] The correction coefficient refers to the operational pressure correction value calculated based on the passenger flow density of the passage, time period attributes (peak / off-peak), and passage type (standard / wide passage) (usually ≤1.0, the value decreases as the pressure increases); the operational pressure quantification model refers to the preset mapping rules or lookup table mechanism used to convert passenger flow, time period, and passage type into pressure coefficients; the multi-source risk decision function refers to the algorithm function that integrates evidence support, anomaly index, and correction coefficient; the fusion strategy refers to the preset weighting rules or Bayesian fusion rules used to integrate the three types of inputs; the comprehensive judgment value refers to the final output, which represents the quantitative result (range 0-1) of the passage risk and credibility.

[0189] The necessary procedures are as follows:

[0190] Correction coefficient calculation: The industrial control all-in-one computer acquires the passenger flow density of the passage (e.g., from step S2 or real-time counting by infrared sensors), the current time period attribute (determined by the system clock as peak / off-peak / nighttime), and the passage type (standard passage or wide passage). These three factors are input into a preset operational pressure quantification model (e.g., a pressure level mapping table or a linear weighted formula) to calculate the correction coefficient. For example, high-density passenger flow (>30 people / minute) combined with peak hours and a standard passage type is classified as high operational pressure, with an output correction coefficient of 0.8; low-density passenger flow (<10 people / minute) combined with off-peak hours is classified as low pressure, with an output correction coefficient of 1.0.

[0191] Multi-source fusion judgment: The decision-making layer simultaneously inputs the evidence support (verification credibility output in step S42), the anomaly index (posture risk value output in step S43), and the correction coefficient (operational pressure value calculated in this step) into a preset multi-source risk decision function. This function performs fusion calculations on the three types of parameters according to a preset fusion strategy (such as weighted product, Bayesian posterior fusion, or DS evidence synthesis rules). For example, when using a weighted strategy, the comprehensive judgment value = evidence support × weight 1 + (1 - anomaly index) × weight 2 × correction coefficient. The final output comprehensive judgment value (range 0-1, higher values ​​indicate higher credibility) is used for subsequent comparison with a preset threshold to generate the final pass judgment result.

[0192] The generation of a uniform level of evidence support includes:

[0193] Step S421: Calculate the environmental disturbance index based on the scene label that generates the verification result and the current passenger flow density of the channel, and call the defensive benchmark credibility from the preset scene-situation credibility mapping table according to the environmental disturbance index. The defensive benchmark credibility is configured to be positively correlated with the environmental disturbance index.

[0194] Among them, the environmental disturbance index refers to the quantitative indicator of the uncertainty of the comprehensive scenario complexity and passenger flow density; the scenario-situation credibility mapping table refers to the preset benchmark credibility database indexed by the environmental disturbance level; the defensive benchmark credibility refers to the credibility threshold reflecting the current environmental risk baseline, which is configured to be positively correlated with the environmental disturbance index (the more complex the environment, the higher the benchmark credibility and the stronger the defense).

[0195] The calculation method employs a two-factor fusion approach: scene labels characterize environmental complexity (e.g., "low light peak" indicates high complexity), and passenger flow density in the passageway characterizes operational pressure. The two factors are weighted to calculate the environmental disturbance index, which is then mapped to a defensive benchmark credibility. The necessary process is as follows: The industrial control unit first extracts environmental complexity parameters (light intensity, occlusion, humidity, etc.) based on the scene labels that generate the verification results (e.g., generated in step S2). Combined with the current passenger flow density in the passageway (from step S2 or real-time counting), the environmental disturbance index is calculated using a weighted formula or lookup table method (e.g., high light intensity + high density = high disturbance index).

[0196] Subsequently, using this index as an index, the corresponding defensive baseline confidence level is retrieved from a pre-defined scenario-situation confidence mapping table. The mapping table pre-stores baseline values ​​according to the disturbance level, and this value is positively correlated with the environmental disturbance index: the higher the environmental disturbance (e.g., "low light peak"), the higher the defensive baseline confidence level (e.g., 0.85), indicating that a higher verification confidence threshold is required; the lower the environmental disturbance (e.g., "normal sparsity"), the lower the baseline confidence level (e.g., 0.70), indicating that standard verification is acceptable.

[0197] Step S422: Identify the verification type of the verification result. The verification type includes electronic credential verification or biometric verification. Based on the identified verification type, call the preset vector construction rule corresponding to the verification type to parse the generation process of the verification result and construct a multi-dimensional credibility vector that reflects the key credibility factors of the verification type.

[0198] In this step, the verification type refers to the category identifier that distinguishes between electronic credential verification and biometric verification (including face, palm vein, or bimodal fusion); the vector construction rule refers to the preset credibility dimension extraction specification defined according to the verification type, which is stored in the rule base; the verification result generation process refers to the flow of key parameters and intermediate states during verification execution (such as biometric extraction quality, matching similarity calculation process, ticket status query results, etc.); the multidimensional credibility vector refers to the numerical vector (each component ranges from 0 to 1) formed by deconstructing the verification process into multiple credibility factor dimensions, which is used to quantify the reliability of verification.

[0199] The necessary procedures are as follows:

[0200] The industrial control all-in-one machine first identifies the verification type: based on the verification result flag output in step S3, it determines whether it is electronic certificate verification, facial feature verification, palm vein feature verification, or dual-modal fusion verification.

[0201] Then the preset vector construction rules are invoked:

[0202] For electronic credential verification, the credibility dimensions defined by the rules include: ticket status credibility (normal / lost / expired), spatiotemporal compliance credibility (location and time period legality), and cross-verification similarity (comparison results with the bound biometric features).

[0203] For biometric verification (face / palm vein / fusion), the credibility dimensions defined by the rules include: image quality credibility (signal-to-noise ratio, sharpness score), feature matching similarity (comparison results with sub-templates), liveness detection credibility (anti-spoofing attack detection results), and scene adaptation credibility (matching degree between environmental parameters and optimization effect).

[0204] Next, the generation process of the verification results is analyzed: the verification execution chain is traced back to extract the original parameters of each dimension. For example, for biometric verification, the quality score after image preprocessing, the similarity calculation value after feature extraction, and the confidence output of the liveness detection algorithm are extracted; for electronic voucher verification, the query results of the entire life cycle of the ticket are extracted, the spatiotemporal rule verification pass flag, and the consistency value calculated by cross-verification are extracted.

[0205] Finally, a multi-dimensional credibility vector is constructed: the extracted credibility values ​​of each dimension (normalized to the 0-1 range) are assembled into a vector form according to the rules defined. For example, electronic certificate verification is constructed as a three-dimensional vector [ticket status credibility, spatiotemporal compliance credibility, cross-verification similarity]; biometric verification is constructed as a four-dimensional vector [image quality credibility, feature matching similarity, liveness detection credibility, scene adaptation credibility]. This multi-dimensional credibility vector is output to step S423 for consistency calibration with the defensive baseline credibility.

[0206] Step S423: The multidimensional credibility vector and the defensive baseline credibility are input into the consistency calibration module in the rule base. The consistency calibration module calculates the credibility deviation between each component of the multidimensional credibility vector and the defensive baseline credibility, and corrects the credibility deviation based on a preset inverse compensation function, thereby outputting a scalarized comprehensive credibility. Here, the consistency calibration module refers to the built-in logic unit in the rule base used to quantify and correct the deviation between the multidimensional credibility and the environmental baseline; the credibility deviation refers to the difference between each component of the multidimensional credibility vector and the defensive baseline credibility; the inverse compensation function refers to a preset algorithm that dynamically adjusts the credibility based on the direction and magnitude of the deviation. Its core logic is: when the credibility is lower than the baseline (negative deviation), it compensates upwards; when it is higher than the baseline (positive deviation), it moderately suppresses or maintains the credibility; the scalarized comprehensive credibility refers to a single numerical credibility (range 0-1) obtained by aggregating the corrected multidimensional information.

[0207] The necessary process is as follows: The industrial control all-in-one computer outputs the multi-dimensional confidence vector V=[v1,v2,…,v] from step S422. n The defensive baseline confidence B input consistency calibration module is invoked in step S421.

[0208] 1. Deviation detection:

[0209] Calculate the confidence deviation Δi between each dimension component and the benchmark: .

[0210] Generate the deviation vector Δ=[Δ1,Δ2,…,Δ n ]. If Δ i <0 indicates that the credibility of this dimension is lower than the defense benchmark.

[0211] 2. Reverse compensation correction:

[0212] The dimensions are corrected based on the preset reverse compensation function f(Δi), and the corrected confidence level vi′ is output:

[0213] In the formula, k1∈(0,1] is the negative deviation compensation coefficient, k2∈(0,1] is the positive deviation suppression coefficient, and B max This represents the upper limit threshold for credibility.

[0214] 3. Scalar Aggregation: The corrected confidence vector V′=[v1′,v2′,…,vn′] is aggregated into a scalarized comprehensive confidence C through a weighted arithmetic mean. Details are as follows:

[0215] The scalar C characterizes the reliability of the verification results after calibration with environmental defensive benchmarks and is output to step S424.

[0216] Step S424: The preset Bayesian posterior fusion algorithm is used to nonlinearly synthesize the comprehensive confidence and the confidence weights assigned by the feature layer to the verification results, thereby outputting a unified evidence support level.

[0217] In this step, the Bayesian posterior fusion algorithm refers to a nonlinear synthesis method that combines prior probabilities and likelihood probabilities based on Bayes' theorem to calculate posterior probabilities; the confidence weight refers to the weight assigned to the verification result based on scene labels in step S41 (as prior probabilities); the comprehensive credibility refers to the scalarized credibility output after consistency calibration in step S423 (as likelihood probabilities); and the unified evidence support refers to the final credibility output after nonlinear synthesis (range 0-1). In the specific implementation of this application, a key parameter of the Bayesian fusion algorithm—namely, "the probability that the system generates high-credibility evidence when the verification fails"—is set to a fixed empirical value (e.g., 0.12). This value is not arbitrarily set, but is a statistical prior value derived from statistical analysis of tens of thousands of confirmed misidentification cases in historical data. This setting allows the Bayesian theoretical framework to be transformed into stable and repeatable computational logic in the specific context of rail transit identity verification.

[0218] The fusion method employs a probability multiplication mechanism: the evidence support is positively correlated with the product of the confidence weight and the overall credibility, but nonlinear compression is achieved through denominator normalization to ensure that a single low-confidence input can significantly lower the final result.

[0219] The fusion process and algorithm implementation are as follows:

[0220] The industrial control integrated machine inputs the confidence weight W (as the prior probability P(H)) assigned in step S41 and the comprehensive confidence C (as the likelihood probability P(E|H)) output in step S423 into the Bayesian posterior fusion module.

[0221] The uniform evidence support S is calculated using the following Bayesian posterior probability formula:

[0222] Among them, molecules The joint probability of a valid and reliable verification is represented by the total probability (which includes a weighted average of the pass and fail cases).

[0223] The algorithm achieves nonlinear synthesis: when both W and C are greater than 0.5, S is significantly higher than the arithmetic mean of the two; when either is lower than 0.5, S is suppressed to be lower than the arithmetic mean of the two, reflecting the security logic of "double confirmation".

[0224] The final output is a unified evidence support level S (range 0-1), which is used by the decision-making layer in step S44 for multi-source risk fusion.

[0225] The multi-source risk decision function, based on a preset fusion strategy, fuses and calculates the input evidence support, anomaly index, and correction coefficient, outputting a comprehensive judgment value including:

[0226] Step S441: Based on the preset basic probability allocation mapping relationship, the evidence support, anomaly index and correction coefficient are mapped to the corresponding basic probability allocation functions.

[0227] In this step, Basic Probability Assignment (BPA) refers to the confidence assignment of each subset of the identification framework (here, {pass, fail}) in DS evidence theory, satisfying the normalization condition (total confidence is 1). The Basic Probability Assignment function refers to the mapping rule that converts input parameters into BPA. The pre-defined Basic Probability Assignment mapping relationship refers to the conversion specifications defined for the three types of heterogeneous inputs (evidence support, anomaly index, and correction coefficient), solving the problems of inconsistent dimensions and differences in physical meaning. The mapping mechanism is achieved through physical meaning alignment and normalization conversion: "high confidence" (evidence support), "low risk" (anomaly index reversed), and "low pressure" (correction coefficient) are uniformly aligned to the high confidence of the "pass" proposition, and vice versa, aligned to the "fail" proposition.

[0228] The detailed mapping process is as follows:

[0229] 1. Evidence Support Mapping (Direct Mapping): The evidence support S (range 0-1, higher values ​​indicate more credible verification) output in step S424 is directly mapped to m1, as follows:

[0230] m1(through) = S × α.

[0231] m1(not passing) = (1 - S) × β.

[0232] m1(Ω) = 1 - m1(pass) - m1(fail). Where α,β∈(0,1] are discount coefficients, reserving a portion of the confidence for the "uncertain" proposition Ω, reflecting the reliability boundary of the evidence.

[0233] 2. Abnormal index mapping (reverse mapping):

[0234] The anomaly index A (range 0-1, higher values ​​indicate more abnormal behavior / higher risk) output in step S43 is inversely mapped to m2 (high anomaly → high confidence level for failure):

[0235] m2(not passing) = A × γ.

[0236] m2(through) = (1 - A) × δ.

[0237] m2(Ω) = 1 - m2(not passing) - m2(passing).

[0238] Where γ,δ∈(0,1] are adjustment coefficients (usually γ=0.95,δ=0.5), ensuring that the "fail" confidence level approaches the upper limit when high-risk behavior (A→1) is achieved, reflecting the defense logic of "abnormality is risk".

[0239] 3. Correction coefficient mapping (discount mapping):

[0240] The correction coefficient R (range 0-1, usually ≤1, such as R=0.8 for high density) output in step S44 is mapped to the environmental pressure discount factor m3:

[0241] m3(through) = R × η.

[0242] m3(not passing) = (1 - R) × θ + ε.

[0243] m3(Ω) = 1 - m3(pass) - m3(fail). Where η and θ are the baseline weights, and ε is the compensation term for misjudgment caused by pressure; or m3 can be directly used as a discount factor in step S443. Finally, three BPA functions m1, m2, and m3 are output, defined on the same identification framework, for evidence synthesis in step S442.

[0244] The mapping design of the basic probability allocation function has a clear operational strategy orientation. For example, the system's pre-set mapping strategy for the "weekday morning rush hour" scenario is to allocate 90% of the evidence support to the "pass" proposition to ensure efficiency, while allocating 95% of the anomaly index to the "fail" proposition to maintain a safety baseline. These mapping coefficients (90%, 95%, etc.) constitute a "scenario-strategy" matrix. The initial value of this matrix is ​​set by expert experience, but it can be periodically fine-tuned through an online learning mechanism based on actual traffic efficiency and safety data in each scenario, enabling continuous optimization of the system's decision-making strategy.

[0245] Step S442: Based on the basic probability allocation function corresponding to the evidence support degree and the anomaly index, the evidence is synthesized according to the preset evidence synthesis rules to obtain the joint basic probability allocation, and the evidence conflict coefficient generated by this synthesis is calculated simultaneously.

[0246] Among them, the evidence synthesis rule refers to the Dempster synthesis rule in DS evidence theory, which is used to perform orthogonal summation on independent evidence sources; the joint basic probability assignment (denoted as m) 12 The composite reliability function is the result of combining m1 (support of evidence) and m2 (abnormality index), defined in the identification framework Θ={pass, fail}; the coefficient of evidence conflict (denoted as K) is a quantitative indicator of the degree of mutual exclusion between two sources of evidence, used to characterize the reliability divergence between the credibility of identity verification and the risk of passage behavior.

[0247] The necessary procedures are as follows:

[0248] The industrial control integrated computer receives m1 (evidence support mapping) and m2 (anomaly index mapping) generated in step S441. First, the evidence conflict coefficient K is calculated: identify mutually exclusive focal element combinations in the two functions (i.e., the intersection of {pass} and {fail}), and calculate the sum of their confidence products: K = m1({pass}) × m2({fail}) + m1({fail}) × m2({pass}).

[0249] Then, joint basic probability assignment is performed. According to Dempster's rule, the following calculations are made for each subset C⊆Θ: .

[0250] Specifically, it includes:

[0251] .

[0252] .

[0253] .

[0254] Output joint fundamental probability assignment m 12 And the conflict coefficient K to step S443.

[0255] Step S443: Based on the basic probability allocation function corresponding to the correction coefficient and the evidence conflict coefficient, the joint basic probability allocation is calibrated through a preset trust adjustment function to obtain the comprehensive basic probability allocation.

[0256] Among them, the trust adjustment function refers to the mapping rule that dynamically adjusts the correction weights based on the conflict coefficient K, used to initiate defensive calibration (prioritizing security) in high-conflict scenarios and perform environmental pressure adaptation (balancing efficiency) in low-conflict scenarios; the comprehensive basic probability allocation (denoted as m) f The final confidence function obtained after calibration is defined as Θ = {pass, fail}, and serves as the direct basis for the decision in step S444.

[0257] The necessary procedures are as follows:

[0258] The industrial control integrated computer receives the joint basic probability allocation m output by step S442. 12 And the conflict coefficient K, and at the same time retrieve the basic probability allocation m3 (environmental pressure discount factor) corresponding to the correction coefficient generated in step S441.

[0259] The system first compares K with a preset conflict threshold κ (usually set to 0.5):

[0260] If K ≤ κ (low conflict): Perform standard synthesis calibration. (The remaining text appears to be incomplete and requires further context.) 12 The composite basic probability assignment m is obtained by combining m3 and m3 according to Dempster's rule. f=m 12 ⊕m3. At this point, the consistency between identity verification and behavior monitoring is relatively good, and the correction coefficient mainly plays a role in regulating environmental pressure (such as appropriately lowering the access threshold during peak hours).

[0261] If K > κ (high conflict): Perform defensive calibration. Calculate the conflict adjustment factor τ = K / (1+K) and apply security enhancement corrections to m12:

[0262] Reduce the reliability of passing: m 12 ′({through})=m 12 ({through})×(1-τ).

[0263] Improve the reliability of {failure}: m 12 ′({not passed})=m 12 ({Not passed})+m 12 ({through})×τ.

[0264] Reliability with uncertainty: m 12 ′(Ω)=m 12 (Ω).

[0265] Then the corrected m 12 The combined output m is obtained by combining m' and m3. f In this case, if the identity and behavior deviate (e.g., verification passes but the posture is abnormal), the system prioritizes security over efficiency and controls risk by increasing the reliability of the {failed} action.

[0266] The final output is the comprehensive basic probability distribution m. f (including m) f ({through}), m f ({Not approved}), m f (Ω))to step S444.

[0267] Step S444: Extract the comprehensive judgment value from the comprehensive basic probability allocation according to the preset decision rules.

[0268] Among them, the decision rule refers to the judgment criterion that maps the basic probability allocation to a binary passage decision based on the confidence comparison; the comprehensive judgment value refers to the scalar value that represents the credibility of passage, and the final decision result is generated by comparing it with a preset threshold.

[0269] The necessary procedures are as follows:

[0270] The integrated basic probability allocation m output by the industrial control all-in-one computer in step S443 is received. f Extract the {pass} reliability and {fail} reliability.

[0271] Calculate the overall decision value D: Take the difference between the {pass} reliability and the {fail} reliability, i.e., D = m f ({through})-mf ({Not approved}).

[0272] Compare D with a preset threshold (usually 0):

[0273] If D≥0, it is determined as passable and passage is allowed;

[0274] If D < 0, it is judged as failing, triggering interception and contextual prompts.

[0275] For example, if m is calibrated after step S443 f ({through}) = 0.4, m f If ({failed}) = 0.5, then D = -0.1 < 0, and the result is deemed a failure; if m f ({through}) = 0.6, m f If ({failed}) = 0.3, then D = 0.3 ≥ 0, and the result is passed.

[0276] Based on the same inventive concept, this invention provides a read / write industrial control integrated machine that implements the aforementioned method. This device is an edge computing device integrating sensing, computing, and control, and it physically implements the access control method of this application through the collaborative work of the following specifically constructed units:

[0277] The data acquisition unit consists of a multi-source sensor hardware group integrated into the device, including a visible light camera for capturing facial images, a near-infrared imaging module for capturing palm vein images, a scanner for reading QR codes, and a reader / writer for sensing contactless tickets. These sensors work in parallel under the scheduling of the device's main control system, providing the system with the passenger's biometric characteristics or original electronic voucher information.

[0278] The conversion unit is implemented by a data standardization service running on the device's main processor. This service calls a pre-built algorithm library to process the raw data in real time, such as normalizing facial images and extracting feature vectors, or parsing QR codes into ticket identifiers. Simultaneously, the service uses onboard environmental sensors (such as light, temperature, and humidity sensors) and a passenger flow counter to acquire environmental parameters and passenger flow density in real time, and synchronizes and encapsulates this contextual information and identity information at millisecond levels, forming a structured data packet with a unified format and scene tags.

[0279] The verification unit is the core processing engine of the device, comprised of the main processor and dedicated processing modules. It contains two parallel verification logic branches:

[0280] 1. Biometric Dynamic Adaptation Verification Branch: After receiving the data packet, this branch generates or matches a scene label based on the environmental parameters contained within. Based on this label, it retrieves an optimized feature extraction parameter set from the device's locally stored preset parameter library and calculates the feature vector using the loaded neural network model. Subsequently, it accesses the feature template library in the device's secure storage area, calls the user sub-template corresponding to the scene label for matching calculation, and outputs the biometric verification result. Internally, it implements the previously described environment adaptive algorithm structure and dynamic update mechanism for the feature template library.

[0281] 2. Electronic Voucher Intelligent Verification Branch: This branch extracts the voucher identifier, queries the backend system's full lifecycle data of the ticket through a secure network channel to determine its status, and performs spatiotemporal compliance verification by combining the device's own location information with the current time period. Simultaneously, this branch triggers a cross-verification process, rapidly comparing any biometric features that may have accompanied the current data collection (such as a face captured during scanning) with the historical biometric features bound to the voucher, and outputs a verification conclusion based on the combined results of all three.

[0282] The fusion decision unit is an intelligent decision-making software system running on the device. This system receives all results from the verification unit and integrates scene labels, real-time passenger flow density in the passageway, and passenger posture data sensed by the infrared sensor array above the device. The system fully implements the three-level weighted algorithm—feature layer, evidence layer, and decision layer—as detailed in the specific implementation: the feature layer assigns weights to evidence based on the scene; the evidence layer fuses and calibrates multi-source identity evidence; and the decision layer quantitatively analyzes behavioral anomalies and operational pressure, generating a comprehensive judgment value through a preset fusion strategy. All algorithm parameters for this unit are derived from the aforementioned large-scale domain data training and optimization process.

[0283] The execution unit consists of the equipment's industrial control interface and drive circuitry. When the overall judgment result is satisfactory, this unit sends a drive signal to the gate via the digital I / O interface to control the gate to open and generates an encrypted transaction log containing verification details for secure storage. When the judgment fails, this unit drives the human-machine interface (such as a display screen or voice module) to output clear and guiding contextual prompts based on the failure reason code.

[0284] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for controlling rail transit traffic, characterized in that, include: The system collects and acquires passengers' biometric or electronic credential information, with the biometrics being collected and acquired through a multimodal identity recognition device deployed on a read / write industrial control all-in-one machine. The collected biometric data or electronic voucher information is converted into identity information in a unified format and synchronously linked to environmental parameters and passenger flow density at the time of collection. Dynamic adaptation verification is performed on identity information. Scene labels are generated based on environmental parameters. If the collected information is biometric information, feature vectors are extracted from the feature extraction parameter set called from the preset parameter library based on the scene label, and matched with the sub-templates of the corresponding scene in the feature template library to output the biometric verification result. If the collected information is electronic voucher information, the voucher identifier is extracted, the ticket status is determined based on the ticket's full life cycle data, and the spatiotemporal compliance is determined by combining the terminal location and time period. The current and bound historical biometric features are cross-verified to determine the degree of consistency, and the verification result of the fusion of ticket status, spatiotemporal compliance and degree of consistency is output. The fusion verification results, combined with scene labels, channel passenger flow density, and passenger posture data perceived by infrared sensor arrays, generate a comprehensive judgment value through a three-level weighted algorithm of feature layer, evidence layer, and decision layer. This comprehensive judgment value is then compared with a preset threshold to generate a final judgment result. Specifically, the feature layer queries a preset feature weight configuration table based on scene labels to assign corresponding confidence weights to the verification results; the evidence layer processes the verification results according to the assigned confidence weights and calls a preset evidence fusion and conflict resolution rule base for comprehensive calculation to generate a unified evidence support level; the decision layer calculates three quantitative indicators—dwelling time anomalies, movement trajectory deviation, and behavioral pattern conflict—based on the passenger posture data perceived by the infrared sensor array, and aggregates them through a preset posture risk assessment model to generate an anomaly index; the decision layer also calculates a correction coefficient based on channel passenger flow density, combined with current time period attributes and channel type, through a preset operational pressure quantification model; finally, the decision layer inputs the evidence support level, anomaly index, and correction coefficient into a preset multi-source risk decision function; this function, according to a preset fusion strategy, fuses the input evidence support level, anomaly index, and correction coefficient, and outputs a comprehensive judgment value. If the overall judgment result is "pass", the gate will be opened and the encrypted transaction log will be recorded. If the overall judgment result is "not passed", a scenario-based prompt will be output.

2. The rail transit traffic control method according to claim 1, characterized in that, If the biometric information is facial features, feature vectors are extracted from a preset parameter library based on scene labels, and then matched with sub-templates of the corresponding scene in the feature template library. The output biometric verification results include: Based on scene labels, the corresponding feature extraction parameter set is called from the preset parameter library; The feature extraction parameter set is optimized using an environment adaptive algorithm; Based on the optimized feature extraction parameter set, extract the feature vector of the face image; Retrieve the face feature sub-template corresponding to the scene label from the feature template library; The extracted feature vectors are matched with the called sub-templates to calculate the similarity value; The matching threshold is dynamically determined based on the scene label. The similarity value is compared with the matching threshold, and the face feature verification result is output.

3. The rail transit traffic control method according to claim 1, characterized in that, If the biometric information is palm vein features, the feature vector is extracted from the feature extraction parameter set called from the preset parameter library based on the scene label, and matched with the corresponding scene sub-template in the feature template library. The output biometric verification results include: Based on scene labels, the corresponding feature extraction parameter set is called from the preset parameter library; Based on the called feature extraction parameter set, extract the feature vector of the palm vein image; Retrieve the palm vein feature sub-template corresponding to the scene label from the feature template library; The extracted feature vectors are matched with the called sub-templates to calculate the similarity value; The matching threshold is dynamically determined based on the scene label. The similarity value is compared with the matching threshold, and the palm vein feature verification result is output.

4. The rail transit traffic control method according to claim 1, characterized in that, If the biometric information includes palm vein features and facial features, feature vectors are extracted from a preset parameter library based on scene labels, and matched with sub-templates of the corresponding scene in the feature template library. The output biometric verification results include: Joint scene labels are generated based on environmental parameters, which include a first type of parameters related to facial features and a second type of parameters related to palm vein features. Based on the joint scene label, the facial feature extraction parameter set and the palm vein feature extraction parameter set are called from the preset parameter library respectively; The parameter sets for facial feature extraction and palm vein feature extraction were optimized using an environment adaptive algorithm. The feature vector of the face image is extracted based on the optimized face feature extraction parameter set, the feature vector of the palm vein image is extracted based on the optimized palm vein feature extraction parameter set, and the feature vectors of the face image and the palm vein image are fused to obtain a joint feature vector. Retrieve the face feature sub-template and palm vein feature sub-template corresponding to the joint scene label from the feature template library; The joint feature vector is matched with the called face feature sub-template and palm vein feature sub-template to obtain a comprehensive similarity value; The matching threshold is dynamically determined based on the joint scene label. The comprehensive similarity value is compared with the matching threshold, and the biometric verification result is output.

5. A method for controlling rail transit traffic according to claim 4, characterized in that, The facial feature extraction parameter set and the palm vein feature extraction parameter set were optimized using an environment adaptive algorithm, including: Obtain the current working mode of the read / write industrial control all-in-one machine, as well as the environmental parameters in the combined scene tag; Based on the current working mode, the facial reference parameter set and palm vein reference parameter set are obtained from the preset mode-parameter mapping relationship respectively; Based on environmental parameters, dynamic parameter sets for faces and dynamic parameter sets for palm veins are generated respectively through preset environmental feature adaptation functions. According to the preset priority rules, the face baseline parameter set and the face dynamic parameter set are weighted and fused to obtain the optimized face feature extraction parameter set, and the palm vein baseline parameter set and the palm vein dynamic parameter set are weighted and fused to obtain the optimized palm vein feature extraction parameter set.

6. The rail transit traffic control method according to claim 1, characterized in that, The generation of a uniform level of evidence support includes: The environmental disturbance index is calculated based on the scene label that generates the verification result and the current passenger flow density of the channel. Based on the environmental disturbance index, the defensive baseline credibility is called from the preset scene-situation credibility mapping table. The defensive baseline credibility is configured to be positively correlated with the environmental disturbance index. The verification result is identified as a verification type, including electronic credential verification or biometric verification. Based on the identified verification type, a preset vector construction rule corresponding to the verification type is invoked to parse the generation process of the verification result and construct a multi-dimensional credibility vector that reflects the key credibility factors of the verification type. The multidimensional credibility vector and the defensive baseline credibility are input together into the consistency calibration module in the rule base. The consistency calibration module calculates the credibility deviation between each dimension component in the multidimensional credibility vector and the defensive baseline credibility, and corrects the credibility deviation based on the preset reverse compensation function, thereby outputting a scalarized comprehensive credibility. A pre-defined Bayesian posterior fusion algorithm is used to non-linearly synthesize the overall credibility and the confidence weights assigned by the feature layer to the verification results, thereby outputting a unified evidence support level.

7. The rail transit traffic control method according to claim 1, characterized in that, The multi-source risk decision function, based on a preset fusion strategy, fuses and calculates the input evidence support, anomaly index, and correction coefficient, outputting a comprehensive judgment value including: Based on the preset basic probability allocation mapping relationship, the evidence support, anomaly index and correction coefficient are respectively mapped to the corresponding basic probability allocation functions; Based on the basic probability allocation function corresponding to the evidence support degree and the anomaly index, the evidence is synthesized according to the preset evidence synthesis rules to obtain the joint basic probability allocation, and the evidence conflict coefficient generated by this synthesis is calculated simultaneously. Based on the basic probability allocation function corresponding to the correction coefficient and the evidence conflict coefficient, the joint basic probability allocation is calibrated by a preset trust adjustment function to obtain the comprehensive basic probability allocation. Based on the preset decision rules, a comprehensive judgment value is extracted from the comprehensive basic probability allocation.

8. A read / write integrated industrial control machine, employing a rail transit traffic control method as described in any one of claims 1-7, characterized in that, include: The data collection unit is used to collect and obtain passengers' biometric or electronic credential information. The conversion unit is used to convert the collected biometric data or electronic voucher information into identity information in a unified format, and synchronously associate it with environmental parameters and passenger flow density at the time of collection. The verification unit is used to perform dynamic adaptation verification on identity information. It generates scene labels based on environmental parameters. If the collected information is biometric information, it extracts feature vectors from the feature extraction parameter set called from the preset parameter library based on the scene label, and matches them with the sub-templates of the corresponding scene in the feature template library, outputting the biometric verification result. If the collected information is electronic voucher information, it extracts the voucher identifier, determines the ticket status based on the ticket's full life cycle data, determines spatiotemporal compliance by combining the terminal location and time period, triggers cross-verification of current and bound historical biometrics to determine the degree of consistency, and outputs the verification result that integrates the ticket status, spatiotemporal compliance, and degree of consistency. The fusion decision unit is used to fuse verification results. It combines scene labels, channel passenger flow density and passenger passage posture data perceived by infrared sensor arrays, and generates a comprehensive judgment value through a three-level weight algorithm of feature layer-evidence layer-decision layer. It then compares the comprehensive judgment result with a preset threshold. The execution unit is used to drive the gate to open and record the encrypted transaction log when the comprehensive judgment result is "pass", and to output a scenario-based prompt when the comprehensive judgment result is "fail".

Images