A target identity self-adaptive verification method in a dynamic environment

Abstract

The application discloses a target identity adaptive verification method in a dynamic environment, relates to the technical field of identity recognition and authentication in the intelligent transportation and smart city related industries, and can relieve the instability of target features caused by environmental condition changes when the existing identity verification scheme performs automatic verification of target identity. The application comprises the following steps: acquiring environment information and reference environment information of a target object, and generating a fusion environment vector; extracting a real-time feature vector and a reference feature vector of the target object, and calculating a similarity degree; inputting the fusion environment vector into a weight generation network to obtain a dynamic weight vector; and performing identity judgment of the target object according to the calculated similarity degree and the dynamic weight vector. The application is suitable for identity recognition and authentication in the intelligent transportation and smart city related industries.

Description

Technical Field

[0001] This invention relates to the field of identity recognition and authentication technology in intelligent transportation and smart city related industries, and in particular to a target identity adaptive verification method in a dynamic environment. Background Technology

[0002] Identity authentication technology is a crucial foundation for ensuring information system security and enabling personalized services. With the rapid development of the Internet of Things (IoT), mobile internet, and artificial intelligence (AI) technologies, identity authentication scenarios have expanded from controlled indoor environments to various complex and dynamic outdoor and mobile scenarios, such as intelligent security monitoring, drone inspections, mobile terminal unlocking, and smart city management. In these real-world applications, the dynamic changes in environmental conditions (such as light intensity, weather conditions, target distance, obstructions, and background noise) pose significant challenges to the accuracy and robustness of identity verification.

[0003] In existing identity authentication schemes, the first step is to acquire the feature information of the target to be verified through a data acquisition device and automatically identify its identity. Then, based on the identified identity, the system retrieves the target features corresponding to that identity from a pre-registered target information database. Finally, a fixed comparison algorithm is used to determine whether the current target matches a target registered in the database. If the similarity is below a set threshold, an identity verification alarm is triggered.

[0004] In recent years, existing solutions have begun to face serious challenges in increasingly complex real-world scenarios, particularly in intelligent transportation and smart cities. This is mainly because existing solutions employ static, single comparison strategies, which are ill-suited to the complex and ever-changing outdoor environments. For example, outdoor environmental interference leads to high false alarm rates. The biometric appearance of a target can vary significantly throughout the day, depending on the time of day, weather conditions, and the equipment used for data collection. Fixed comparison algorithms are highly susceptible to misjudgments due to these legitimate changes in appearance (such as changes in lighting angle or shadow occlusion), resulting in high false alarm rates and disrupting normal identity authentication processes.

[0005] Furthermore, because existing solutions primarily process the characteristics of the target itself, they neglect the crucial contextual information of "the environmental conditions under which the target was collected." The specific recognition algorithm cannot distinguish between situations where "backlighting reduces target contrast" and "differences in the target's inherent features," leading to data errors caused by environmental changes being incorrectly interpreted as target identity mismatch. There is also the problem of rigid comparison strategies. For example, regardless of whether it's foggy, nighttime, or a strong backlighting scene, the algorithm compares all feature parts of the target in the same way. In reality, shape information may be more reliable under low light conditions, while color information is more stable under sufficient light. In other words, there is a clear "chain" relationship between environmental detection and authentication decisions; environmental changes trigger parameter adjustments, but the threshold adjustment strategy is relatively fixed, resulting in a pre-set, rule-driven adjustment strategy that lacks online adaptive optimization capabilities for the identity verification model itself. When environmental changes exceed the coverage of pre-set rules, system performance may still experience a precipitous drop. The perception granularity of physical environmental factors (such as light, distance, and obstructions) is too coarse, failing to fully explore the deep coupling relationship between target features and environmental features. This results in frequent false alarms and missed detections due to the rigidity of the strategy and its inability to follow environmental changes.

[0006] Therefore, how to further improve the identity recognition and authentication technology applied in outdoor scenarios in the intelligent transportation and smart city industries, and alleviate the problem of high false alarm rate and high false negative rate caused by the instability of target characteristics due to changes in environmental conditions when the existing identity verification scheme performs automatic verification of target identity, has become a key research topic. Summary of the Invention

[0007] The embodiments of the present invention provide a target identity adaptive verification method in a dynamic environment, which can alleviate the problem of high false alarm rate and high false negative rate caused by the instability of target features due to changes in environmental conditions when existing identity verification schemes perform automatic target identity verification in outdoor environments.

[0008] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: An adaptive verification method for target identity in a dynamic environment, the method being used in an identity verification system for an outdoor environment, the method comprising: S1. Obtain the environmental information and baseline environmental information of the target object, and generate a fused environmental vector, wherein the fused environmental vector is obtained from the original real-time environmental vector and the original baseline environmental vector; S2. Extract the real-time feature vector and the baseline feature vector of the target object, and calculate the similarity degree, wherein the target image of the target object is divided into component regions, and the similarity degree includes the numerical value of the similarity degree corresponding to each component region; S3. Input the fusion environment vector into the weight generation network to obtain the dynamic weight vector; S4. Based on the calculated similarity and the dynamic weight vector, determine the identity of the target object.

[0009] In this embodiment, S1 includes: when the acquisition device captures a target image of the target object, triggering the environmental information acquisition process and obtaining the original real-time environmental vector. The environmental information acquisition process includes: acquiring spatiotemporal information records and acquiring meteorological API data based on the spatiotemporal information records; querying the reference image of the target object from the archive database based on the target image, and obtaining environmental information corresponding to the reference image, and constructing an original reference environmental vector. ;right and The continuous numerical variables in the data are standardized, with special encoding for continuous variables with periodic characteristics; and a fusion environment vector is generated based on the standardization results and the absolute difference in dimensions.

[0010] The standardization process includes mapping each dimension of the continuous numerical variable to a uniform statistical scale. Specifically, for the j-th dimension, the standardization is performed as follows: the standardized output value of the j-th dimension... ,in, For the j-th dimension, the original continuous numerical value. To train the average value of the j-th dimension in the dataset, The standard deviation of the j-th dimension in the training dataset; the standardization results include: d represents the total number of environmental dimensions. For the standardization of current environmental characterization, Standardized characterization of the basic environment This represents the value of the d-th dimension in the current environment. This represents the value of the d-th dimension in the base environment, and T represents the transpose operation.

[0011] Special encoding is applied to continuous variables exhibiting periodic characteristics, including: converting continuous variables with periodic characteristics into a two-dimensional continuous representation, using the following encoding method: t represents the time variable, and T represents the transpose operation. This represents the time period encoding vector.

[0012] Based on the standardization results and the absolute differences in dimensions, a fusion environment vector is generated, including: the absolute differences in dimensions represented as a difference vector. ; Fusion environment vector , This indicates a vector concatenation operation.

[0013] In this embodiment, S2 includes: dividing the target image into regions, and then extracting features to obtain a real-time feature vector. and baseline eigenvectors ; obtain and The similarity is vectorized to obtain the degree of similarity.

[0014] Among them, obtaining and The similarity is vectorized to obtain the similarity degree, including: obtaining the feature vector of component region i. The target image is divided into component regions i, i=1,2,...,N. , It is the C-dimensional feature vector of the k-th layer feature map at position (x, y), where x and y are the X-axis and Y-axis coordinates, respectively. The k-th layer of the backbone network serves as the feature extraction layer for the component region. It is the region on the feature map that component region i is mapped to. It is the number of pixels within component region i; the number of pixels in component region i and The similarity is , This represents the feature vector extracted from the i-th component region in the real-time captured image. Let represent the feature vector extracted from the reference image for the i-th component region, where C represents the maximum dimension of the feature vector (i.e., the number of channels), i represents the component number, and j represents the j-th element in the feature vector. The number of elements in the feature vector corresponds to the number of dimensions. This represents the element of dimension j corresponding to the feature vector extracted from the i-th component region in the real-time captured image. Let represent the element of dimension j corresponding to the feature vector extracted from the reference image for the i-th component region, and C represent the maximum dimension of the feature vector; the similarity of the N component regions is arranged in the order of component region numbers and constitutes an N-dimensional difference vector. .

[0015] In this embodiment, S3 includes: establishing a weight generation network, wherein the first hidden layer of the weight generation network is a feature abstraction layer, the second hidden layer is a cross-modulation layer, the environment vector is fused into the weight generation network, and the original score vector z with the same dimension as the component region is output. , Let represent the set of real numbers; wherein, during the training process of the weight generation network, the weighted matching score corresponding to the dynamic weight vector is imported into the loss function. Weight normalization is then performed to obtain the dynamic weight vector.

[0016] Among them, the first hidden layer Second hidden layer , and These are the trainable parameters for the first hidden layer. and These are the trainable parameters for the second hidden layer. This is the first hidden layer. This is the second hidden layer. The activation function is ; the loss function is . Where S is the weighted matching score, , Let w represent the dynamic weight vector of component region i, and w represent the normalized weight vector. , and These represent two boundary values ​​related to environmental change. This indicates the current sample's identity label.

[0017] In this embodiment, S4 includes: weighting and aggregating the component difference vector based on the dynamic weight vector to obtain an identity verification score, comparing the verification score with a preset decision threshold, and determining the target identity based on the comparison result.

[0018] The adaptive target identity verification method in a dynamic environment provided by this invention acquires real-time target features and environmental data, retrieves baseline target features and environmental data, fuses environmental data to generate a fused environmental vector, extracts target component features and calculates differences, dynamically generates verification weights for each component based on the fused environmental vector, and uses the weights to weight the component differences to obtain the final verification result. This implements an "environment-verification" linkage mechanism for identity verification, treating real-time environmental data as an independent input, processing it in parallel with target image data, and deeply fusing them at the decision level to achieve highly robust identity verification. This alleviates the problem of high false positive and high false negative rates caused by unstable target features due to changes in environmental conditions when performing automatic target identity verification in existing identity verification schemes. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention;

[0021] Figure 2 A flowchart of the target identity verification process provided in an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Embodiments of the present invention will be described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of the present invention means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0023] This invention aims to address the high false positive and false negative rates in existing identity verification systems when performing automatic identity verification due to unstable target features caused by changing environmental conditions. Specifically, it provides a method and apparatus capable of synchronously sensing real-time environmental information of the target and dynamically adjusting the verification strategy accordingly. This enables the system to intelligently distinguish between "legitimate appearance differences caused by environmental factors" and "essential differences caused by illegitimate identity," thereby significantly improving the accuracy and reliability of verification. The general design concept involves introducing an environmental sensing module into the identity verification system to make its verification decisions more accurate. When comparing target features, the system simultaneously analyzes the current environmental conditions and dynamically determines "which part of the target's features should be emphasized" as the basis for the final judgment.

[0024] The method designed in the embodiments of the present invention, such as Figure 1 As shown, its main process includes:

[0025] S1. Obtain the environmental information and baseline environmental information of the target object, and generate a fused environmental vector;

[0026] S2. Extract the real-time feature vector and baseline feature vector of the target object, and calculate the similarity.

[0027] S3. Input the fusion environment vector into the weight generation network to obtain the dynamic weight vector;

[0028] S4. Based on the calculated similarity and the dynamic weight vector, determine the identity of the target object.

[0029] In the specific implementation process, in order to realize the above method and process, it is necessary to build a corresponding adaptive verification system. From the perspective of system architecture construction and program design, it can be divided into four parts, including: environmental data synchronization and fusion, target feature extraction and difference calculation, dynamic verification weight generation and adaptive weighted verification decision.

[0030] Part 1: Environmental Data Synchronization and Fusion. The purpose of this part is to quantitatively fuse information from real-time acquired environmental data with that of a baseline environment, generating a numerical vector that characterizes the challenges posed by environmental differences. This provides the input basis for subsequent dynamic weight generation, specifically including:

[0031] Step 1.1: Real-time Continuous Acquisition of Environmental Data: When the acquisition device completes the capture of the target, the system synchronously triggers the environmental data acquisition process. This includes recording the spatiotemporal information and meteorological API data requests, recording the timestamp T (accurate to the second) of the capture time and the geographic coordinates (Lon, Lat) of the capture location. This spatiotemporal information will be used to initiate a query to the meteorological data service interface. The adaptive verification system sends a request to the meteorological service API, with request parameters including at least the timestamp T and coordinates (Lon, Lat). The API returns the real-time environmental data for that location and time, forming the original real-time environmental vector: Each component e j These are continuous numerical variables, including but not limited to: light intensity, precipitation, visibility, relative humidity, temperature, wind speed, atmospheric pressure, etc., j=1,2,3…m, where m is the total number of component types. These variables together constitute a continuous description of the physical environment at that moment, and the vectorized expression of the component types is called the environmental dimension (or simply dimension).

[0032] Step 1.2, Acquisition of baseline environment data: Based on the captured target image, the adaptive verification system queries the baseline image of the target from the archive database and simultaneously acquires the environmental information associated with the baseline image to form the original baseline environment vector. The processing principle is as follows: If the reference image records real environmental data during registration: then Also a continuous variable vector obtained from a meteorological API, and They have identical dimensions and units. If the baseline image does not record environmental information: a standard baseline environment is set as the default value. The default value is the statistical median of each environmental dimension in the training dataset. Using the median instead of the mean avoids excessive influence of extreme weather samples on the default value.

[0033] Step 1.3, Standardization of Continuous Variables: Since environmental variables have different physical dimensions and numerical ranges, direct use can lead to difficulties in subsequent network training. Therefore, independent Z-score standardization is performed on each dimension to map them to a uniform statistical scale. For the j-th environmental dimension, the standardization formula is: ,in These are the original continuous values ​​for this dimension. To train the average value of this dimension in the dataset, The standard deviation of this dimension in the training dataset.

[0034] After standardization, two standardized vectors are obtained: Where d is the total number of environmental dimensions, For the standardization of current environmental characterization, It is a standardized representation of the basic environment.

[0035] Step 1.4: Special Encoding for Periodic Variables: For continuous variables with periodic characteristics, directly using the original numerical values ​​will lose the information about the periodic boundaries. For such variables, the implementation method of this invention uses sine-cosine encoding to convert them into a two-dimensional continuous representation: This encoding method ensures the periodic continuity of the time variable, meaning that the Euclidean distance between any two time points in the sine-cosine space is proportional to their actual time difference. After this encoding, the time dimension is expanded from one dimension to two dimensions and serves as the environment vector. and One of the components.

[0036] Step 1.5, Environmental Difference Quantification and Vector Fusion: Calculate the absolute difference for each dimension to form a difference vector. Then, the three vectors are concatenated along the dimensional direction to form a fused environment vector. ,in This indicates a vector concatenation operation.

[0037] Step 1.6, Output and Subsequent Interface: The final generated fused environment vector will be used as input and fed into the environment-aware weight generation network described in the subsequent steps of this invention. This network will dynamically calculate the weights used for identity verification decisions based on this vector. It should be noted that the above fusion method is a preferred embodiment of this invention. Other alternative fusion strategies can also be used without departing from the core concept of this invention, such as using only difference vectors, weighted concatenation, nonlinear combination, etc.

[0038] Part Two: Target Feature Extraction and Difference Calculation. The purpose of this part is to extract physically meaningful target component features from real-time capture data, quantify the similarity of these component features, and generate a component difference vector. This provides the foundational data for subsequent dynamic weighting, specifically including:

[0039] Step 2.1, Target Region Definition: Before feature extraction, it is necessary to clearly define which components of the target will be included in the subsequent difference calculation. In a preferred embodiment of the present invention, the target is divided into N component regions with different physical meanings and environmental sensitivities, such as: the front region, top contour, left region, right region, bottom region, core identifier region, edge contour region, and auxiliary texture region. The division of these N components is not fixed and can be adjusted according to the actual application scenario. For example, in a scenario where only forward capture is required, only the relevant front region can be retained; in a scenario requiring higher precision, the front region can be further subdivided into upper, middle, and lower parts.

[0040] Step 2.2, Selection and Configuration of Fixed Feature Extraction Model: A pre-trained deep neural network can be used as the backbone model for feature extraction. The parameters of this model are completely fixed in this adaptive verification system and do not participate in subsequent training updates. The preferred backbone model for feature extraction is a residual network (ResNet-50) or its improved version pre-trained on the target image dataset. The input of this backbone model is the target image, and the output is feature maps of multiple intermediate layers. To obtain part-level features, the following method is used: For each predefined part region i (i=1,2,...,N), based on its coordinate position (x1,y1,x2,y2) in the image, a region of interest pooling operation is performed on the feature map of the intermediate layer of the backbone network to obtain a fixed-dimensional feature vector f. Wherein, if the k-th layer of the backbone network is selected as the part region feature extraction layer, the size of the feature map of this layer is (H*W*C). For component region i, it is mapped to the feature map size using bilinear interpolation to obtain the corresponding feature map region. Then, global average pooling is performed on all feature vectors within this region to obtain the feature vector of the component region. , It is the C-dimensional feature vector of the k-th layer feature map at position (x,y). It is the region on the feature map that component region i is mapped to. It represents the number of pixels in that area.

[0041] Step 2.3, Parallel Feature Extraction of Real-Time and Reference Images: For each image to be verified, i.e., the real-time image and the reference image, the above feature extraction process is executed separately to obtain the real-time feature vector of the component region features. With reference eigenvectors .

[0042] Step 2.4, Calculate the similarity of component region features: For component region i, calculate the real-time feature vector. With reference eigenvectors The similarity between them. The preferred similarity metric in this invention is cosine similarity, calculated using the following formula: The cosine similarity value ranges from [−1, 1]. A value closer to 1 indicates that the two feature vectors are more aligned in direction, meaning the two component regions are more similar in appearance; a value closer to -1 indicates opposite directions, meaning greater differences. For some component regions that are sensitive to numerical ranges, other similarity measures can be used as supplements or alternatives, such as ROC recognition comparison, the reciprocal of Euclidean distance, and Pearson correlation coefficient. However, in the preferred embodiment of this invention, cosine similarity is used uniformly to ensure consistency and efficiency in calculation.

[0043] Step 2.5: Construction of the component region dissimilarity vector: Arrange the similarity scores si calculated for each of the N component regions in order of component region number to form an N-dimensional component region dissimilarity vector. This vector fully describes the similarity between the real-time target and the baseline target at the level of each component region.

[0044] Step 2.6, Output and Subsequent Interface: The final generated component region difference vector will be used as one of the inputs, and together with the fusion environment vector generated in the first part, it will be sent to the weighted aggregation module described in the subsequent steps of this invention. This module will perform a weighted summation of the component region differences according to dynamic weights to obtain the final identity verification score.

[0045] It should be noted that the above-mentioned division of the N component regions is a preferred embodiment of the present invention, but not the only one. In practical applications, the definition of component regions can be adjusted according to the following principles, such as the granularity adjustment principle: if more refined discrimination is required, large component regions (such as the "target front region") can be further subdivided into the "front upper region" and the "front lower region"; if computational efficiency is pursued, multiple component regions can be merged (such as merging the "left region" and the "right region" into the "lateral region"); another example is the scene adaptation principle: the definition of component regions can differ for different types of targets (small targets, large targets, irregular targets). For example, a "top structural feature" component region can be added for large targets, and a "prominent structural feature" component region can be added for targets with special structures; yet another example is the dynamic selection principle: in some cases, certain component regions may be occluded or exceed the image boundary (such as the top region of a large target). In this case, these component regions can be dynamically masked, and the similarity can only be calculated for the visible component regions, and the weight vector can be renormalized during subsequent weighting.

[0046] Part Three: Dynamic Verification Weight Generation. The purpose of this part is to dynamically calculate a set of weights based on the fusion environment vector generated in Part One. These weights are used to measure the reliability and importance of each component region in the current environment, providing a basis for subsequent weighted verification decisions. Specifically, this includes:

[0047] Step 3.1, Definition of the Fusion Environment Vector: The input to this step is the fusion environment vector generated in Part 1. As described in Part 1, this vector is formed by concatenating the real-time environment normalization vector, the baseline environment normalization vector, and the difference vector between the two, and its dimension is 3D. This vector fully represents the absolute state of the environment faced by the current verification task and the degree of environmental change from the baseline to the real-time environment.

[0048] Step 3.2, Weight Generation Network Structure Design: Design a lightweight fully connected neural network W as the weight generator, with parameters denoted as... The network is based on As input, the output is an original score vector with the same dimensions as the component region. The network architecture uses a multilayer perceptron, and the specific design is as follows:

[0049] The first hidden layer is a feature abstraction layer. This layer maps the input D-dimensional vector to a higher-dimensional hidden space to extract combined patterns of environmental features. Its calculation formula is as follows: ,in and These are the trainable parameters for the first hidden layer. This is the first hidden layer. This is the activation function.

[0050] To further integrate environmental characteristics and simulate the cross-influence of different environmental factors on the importance of component regions, a second hidden layer is set, which is a cross-modulation layer. This layer can be a linear transform with a gating mechanism or a standard fully connected layer. In a preferred embodiment, a standard fully connected layer with ReLU activation is used. ,in and These are the trainable parameters for the second hidden layer. This is the second hidden layer. This is the activation function.

[0051] The output layer generates the original scores by mapping the output of the second hidden layer to an N-dimensional space, obtaining the original score z for each component region. The standard formula for calculating z is: ,in, and This represents the two types of trainable parameters in the output hidden layer. The specific values ​​of the trainable parameters differ for each part region, and the original score vector... Each vector in It represents the importance of the i-th component region in the current environment, i=1,2,3…N.

[0052] Step 3.3, Weight Normalization: To ensure that the output weight vector satisfies... To address the constraints, this invention employs the Softmax normalization function to convert the original scores into a final dynamic weight vector. ,in This is an optional temperature parameter used to control the smoothness of the weight distribution. When When the value is larger, the weight distribution is more even; when When the value is small, the weight distribution is sharper, tending to highlight a few of the most important component areas. In this invention, It can be set to a fixed value or used as a trainable parameter to be automatically learned by the network. After Softmax normalization, each component of the weight vector w is between 0 and 1, and the sum is 1, which can be directly used as the coefficient for subsequent weighted aggregation.

[0053] Step 3.4, Network Training Process: The weight generation network needs to be jointly trained with the target of the entire adaptive verification system. The training process follows the principle that each training sample is a triplet. ,in Let s be the fused environment vector, y be the component region difference vector, and y be the target unique label. For each sample, perform the following calculation: First, s... Input the weights into the weight generation network to obtain the dynamic weight vector w. Then calculate the weighted matching score S, with the following formula: .

[0054] An adaptive contrastive loss function based on the magnitude of environmental changes is designed. First, from the fused environment vector... Extract the difference part Calculate its L1 paradigm as the total magnitude of its environmental change: Then, positive and negative sample boundaries are generated based on the magnitude of environmental changes. , ,in All are preset hyperparameters. The loss function is defined as: S is the weighted matching score. Boundary value and Adaptability to environmental changes: In harsh environments where features are generally unreliable, the network will degrade. And improve This makes distinguishing tasks relatively easy to achieve; however, in favorable environments, the boundaries become stricter, requiring highly separable scores. The gradient of the loss function L with respect to the parameters θ of the weight generation network and the boundary prediction sub-network is calculated using the backpropagation algorithm. During training, only the parameters of the weight generation network and the boundary prediction sub-network are updated, while the parameters of the backbone feature extraction network remain fixed. The optimizer can employ standard algorithms such as Adam or SGD.

[0055] Step 3.5, Output and Interpretation of Dynamic Weights: After sufficient training, the weight generation network can output and interpret the dynamic weights based on the input fusion environment vector. Output a reasonable weight distribution. This weight distribution has clear physical interpretability. When When the display shows "extremely low light intensity and heavy rainfall" (i.e., light rain at night), the weight values ​​corresponding to "highlighted / reflective areas in front of the target" and "identification areas" in the network's output weight vector w will significantly increase, while the weight values ​​corresponding to "target surface texture and color features" will decrease. This indicates that the adaptive verification system automatically determines that in rainy night environments, the highlighted areas in front and the identification areas are more reliable identification criteria, while surface colors are unreliable. This dynamic weight adjustment mechanism enables the entire adaptive verification system to have environmental awareness capabilities, allowing it to adopt the most appropriate matching strategy for different environmental conditions, thereby significantly improving verification accuracy in complex environments.

[0056] Part Four: Adaptive Weighted Verification Decision. Part Four involves using the component region difference vector *s* and dynamic weight vector *w* generated in the preceding steps to calculate the final environmental adaptive verification score. Based on this score, a determination is made regarding whether the target identity is consistent. Specifically, this includes:

[0057] Step 4.1: The component area difference vector s and the dynamic weight vector w are weighted and summed to obtain the verification score S'. The verification score can be calculated using a weighting method similar to the weighted matching score; these are two scores with similar calculation approaches but different applications. Where N is the total number of component regions. satisfy , The similarity score represents the degree of similarity between the i-th component region and the baseline objective under the current environmental conditions.

[0058] Step 4.2, Threshold Comparison and Judgment: Compare the verification score S' with a preset decision threshold. If it is greater than or equal to the decision threshold, the identity is determined to be consistent; if it is lower than the decision threshold, the identity is determined to be abnormal. The threshold can be determined by optimizing preset indicators on the validation set.

[0059] This invention aims to address the high false positive and false negative rates in existing identity verification systems when performing automatic identity verification due to unstable target characteristics caused by changes in environmental conditions. Specifically, it provides a method and apparatus capable of synchronously sensing real-time environmental information of the target and dynamically adjusting the verification strategy accordingly. This enables the system to intelligently distinguish between "legitimate appearance differences caused by environmental factors" and "essential differences caused by illegitimate identity," thereby significantly improving the accuracy and reliability of verification.

[0060] Specifically, the system acquires real-time target features and environmental data, retrieves baseline target features and environmental data, fuses environmental data to generate a fused environmental vector, extracts target component features and calculates differences, dynamically generates verification weights for each component based on the fused environmental vector, and uses these weights to weight the component differences to obtain the final verification result. This implements an "environment-verification" linkage mechanism for identity verification, using real-time environmental data as an independent input, processing it in parallel with target image data, and deeply fusing them at the decision-making level to achieve highly robust identity verification. This alleviates the high false positive and high false negative rates caused by unstable target features due to changing environmental conditions in existing identity verification schemes during automatic target identity verification. Simultaneously, a network specifically generating verification weight vectors, driven by the "difference between real-time and baseline environments," enables the system to intelligently switch strategies to cope with different environmental challenges.

[0061] For example, in practical applications, this embodiment can understand legitimate differences in target appearance caused by environmental changes such as weather and lighting, avoiding a large number of invalid alarms caused by environmental changes and significantly reducing the verification burden on system operators. It improves detection rates in harsh environments: In scenarios where the performance of traditional algorithms degrades, such as fog, rain, and night, by dynamically focusing on the most reliable features at the time (such as highlighted areas and outlines), the target identity verification capability is enhanced, thereby reducing the false negative rate. It enhances system intelligence and credibility: The verification process is no longer a black box; the generated weight vector provides a basis for decision-making, enabling managers to understand why the system makes a certain judgment under specific weather conditions, increasing the system's acceptability and maintainability.

[0062] This invention is applicable to scenarios such as smart security, IoT device authentication, and mobile terminal identity verification. It can dynamically adjust the identity verification strategy, model, or parameters according to changes in environmental conditions to improve the accuracy and robustness of identity recognition.

[0063] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The above descriptions are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A target identity adaptive verification method in a dynamic environment, characterized in that, The method is used in an identity verification system for an outdoor environment, and the method includes: S1. Obtain the environmental information and baseline environmental information of the target object, and generate a fused environmental vector, wherein the fused environmental vector is obtained from the original real-time environmental vector and the original baseline environmental vector; S2. Extract the real-time feature vector and the baseline feature vector of the target object, and calculate the similarity degree, wherein the target image of the target object is divided into component regions, and the similarity degree includes the numerical value of the similarity degree corresponding to each component region; S3. Input the fusion environment vector into the weight generation network to obtain the dynamic weight vector; S4. Based on the calculated similarity and the dynamic weight vector, determine the identity of the target object; S3 includes: establishing a weight generation network, wherein the first hidden layer of the weight generation network is a feature abstraction layer, the second hidden layer is a cross-modulation layer, the environment vector is fused into the weight generation network, and the original score vector z with the same dimension as the component region is output. , Let represent the set of real numbers; wherein, during the training process of the weight generation network, the weighted matching score corresponding to the dynamic weight vector is imported into the loss function; Then, weight normalization is performed to obtain the dynamic weight vector; Among them, the first hidden layer Second hidden layer , and These are the trainable parameters for the first hidden layer. and These are the trainable parameters for the second hidden layer. This is the first hidden layer. This is the second hidden layer. The activation function is ; the loss function is . Where S is the weighted matching score, , Let w represent the dynamic weight vector of component region i, and w represent the normalized weight vector. , and These represent two boundary values ​​related to environmental change. This indicates the current sample's identity label.

2. The method according to claim 1, characterized in that, S1 includes: When the acquisition device captures an image of the target object, it triggers the environmental information acquisition process and obtains the original real-time environmental vector. The environmental information acquisition process includes: acquiring spatiotemporal information records and acquiring meteorological API data based on the spatiotemporal information records; Based on the target image, the reference image of the target object is retrieved from the archive database, along with the environmental information corresponding to the reference image, and an original reference environment vector is constructed. ; right and The continuous numerical variables in the data are standardized, and continuous variables with periodic characteristics are encoded. Based on the standardization results and the absolute difference in dimensions, a fusion environment vector is generated.

3. The method according to claim 2, characterized in that, The standardization process includes: For continuous numerical variables, each dimension is mapped to a uniform statistical scale. For the j-th dimension, the standardization process is as follows: the standardized output value of the j-th dimension... , For the j-th dimension, the original continuous numerical value. To train the average value of the j-th dimension in the dataset, The standard deviation of the j-th dimension in the training dataset; The standardized processing results include: d represents the total number of environmental dimensions. For the standardization of current environmental characterization, Standardized characterization of the basic environment This represents the value of the d-th dimension in the current environment. This represents the value of the d-th dimension in the base environment, and T represents the transpose operation.

4. The method according to claim 2, characterized in that, Encoding continuous variables with periodic characteristics includes: Continuous variables with periodic characteristics are encoded and converted into a two-dimensional continuous representation. The encoding method is as follows: t represents the time variable, and T represents the transpose operation. This represents the time period encoding vector.

5. The method according to claim 3, characterized in that, Based on the standardization results and the absolute difference in dimensions, a fusion environment vector is generated, including: The absolute difference of dimensions is represented as a difference vector. ; Fusion environment vector , This indicates a vector concatenation operation.

6. The method according to claim 3, characterized in that, S2 include: The target image is divided into regions, and then features are extracted to obtain real-time feature vectors. and baseline eigenvectors ; Get and The similarity is vectorized to obtain the degree of similarity.

7. The method according to claim 6, characterized in that, Get and The similarity is vectorized to obtain the degree of similarity, including: Obtain the feature vector of component region i Wherein, the target image is divided into component regions i, i=1,2,...,N, and N is the total number of component regions divided into the target image. , It is the C-dimensional feature vector of the k-th layer feature map at position (x, y), where x and y are the X-axis and Y-axis coordinates, respectively. The k-th layer of the backbone network serves as the feature extraction layer for the component region. It is the region on the feature map that component region i is mapped to. It is the number of pixels within component region i; Component area i and The similarity is , This represents the feature vector extracted from the i-th component region in the real-time captured image. This represents the feature vector extracted from the reference image for the i-th component region. This represents the element of dimension j corresponding to the feature vector extracted from the i-th component region in the real-time captured image. Let J represent the element of dimension j of the feature vector extracted from the reference image for the i-th component region, and C represent the maximum number of dimensions of the feature vector. The similarity scores of the N component regions are arranged in order of their component region numbers and form an N-dimensional difference vector to represent the degree of similarity. .

8. The method according to claim 1, characterized in that, S4 include: The identity verification score is obtained by weighting and aggregating the component difference vector based on the dynamic weight vector. The identity verification score is then compared with a preset decision threshold, and the target identity is determined based on the comparison result.

Images