Face recognition security method based on AI

Abstract

The invention relates to the technical field of security and protection monitoring, in particular to an AI-based face recognition security and protection method, which comprises the following steps: recognizing a channel based on monitoring an entrance and exit, collecting and analyzing a multispectral face image, screening feature points and skin color areas, optimizing spatial feature matching and structural distribution, and fusing head actions and face behavior parameters. And judging structural consistency and linkage features, and outputting a face recognition mapping result. According to the method, facial detail level discrimination is realized through multi-spectral information fusion, parallel matching of continuous spatial features and time sequence behavior parameters and synchronous screening of behavior linkage and facial structure features are carried out based on regional pixel distribution and a structure sequence in combination with dynamic spatial trajectory analysis and abnormal motion correction, and therefore, the accuracy and the reliability of the method are improved. According to the multi-parameter joint recognition strategy in the dynamic scene, interference caused by environment change and human intervention is reduced, the reliability of target identity judgment in a complex place is improved, and the practicability and the real-time recognition capability for the multi-source scene are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of security monitoring technology, and in particular to an AI-based face recognition security method. Background Art

[0002] The security monitoring sector involves information collection and processing systems that enable real-time perception, identification, and recording of specific areas, objects, or individuals. These systems primarily include video surveillance, behavioral analysis, image recognition, identity verification, and abnormal event detection. These systems aim to enhance the security and protection capabilities of a facility and are widely used in a variety of scenarios, including public safety, traffic management, community governance, and corporate security. Traditional facial recognition security methods utilize image acquisition equipment to capture a target facial image, extract key facial features through a feature extraction algorithm, and then compare these features with feature templates in a facial database to achieve identity recognition, thereby verifying and recording people entering and exiting a facility.

[0003] Existing technologies are based on the comparison of static feature points under a single spectrum, lack the ability to coordinate the determination of temporal and spatial parameters in dynamic scenes, react slowly to behavioral changes and multi-scene switching, and are unable to effectively respond to continuous changes such as motion status, lighting disturbances, and facial micro-movements. They are unable to achieve adaptive fusion of multi-dimensional features, and are prone to omissions in the identification of abnormal targets, misjudgment of camouflage, and failure of multi-source signal fusion, affecting the adaptation and response of security monitoring to changing locations. Summary of the Invention

[0004] The purpose of this invention is to solve the shortcomings of the existing technology and propose an AI-based face recognition security method.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an AI-based face recognition security method, comprising the following steps: S1: Based on the monitoring entrance and exit recognition channel, analyze the pixel distribution of the facial image, determine the contour clarity captured by the infrared vision array, compare the acquisition details of the multispectral sensor, screen the feature points and symmetry parameters of the complete frame, determine the skin color area distribution, and obtain the multispectral facial features; S2: Based on the multispectral facial features, calling the image recognition buffer unit, optimizing the multi-frame images of the facial feature aggregation channel, judging the uniformity of the spatial distribution of feature points, comparing the overlap of contour structures, screening consecutive facial frames with spatial differences within an interval, and obtaining a spatial feature screening sequence; S3: Based on the spatial feature screening sequence, analyze the three-dimensional spatial changes of feature points, compare the X-axis and Y-axis position trends of consecutive frames, determine the Z-axis trajectory continuity, screen feature points with stable position changes, perform corrections on trajectory mutation points, and obtain a three-dimensional trajectory sequence; S4: Based on the three-dimensional trajectory sequence, calculate the change in head rotation angle, determine the speed of facial micro-movements recorded by the local motion capture unit, compare the movement parameters with the spatial coordinate change pattern of the feature points, screen the data with the best matching degree, and obtain the movement linkage matching data.

[0006] The improvements of the present invention are as follows: the multispectral facial features include spectral distribution parameters, texture feature quantities, and skin color block information; the spatial feature screening sequence includes spatial consistency indicators, structural aggregation labels, and inter-frame stable segments; the three-dimensional trajectory sequence includes coordinate trajectory data, coherence marks, and corrected trajectory sets; and the action linkage matching data includes behavior linkage parameters, action identification numbers, and linkage matching relationship groups.

[0007] The present invention is improved in that the steps of acquiring the multispectral facial features are specifically as follows: S111: Based on the monitoring entrance and exit recognition channel, the pixel distribution of the facial image data is analyzed. For each block of the image sequence, the distribution density of the edge continuity of each pixel block is calculated by statistically analyzing the change level of the grayscale gradient. The edge clarity change trend of each area in the infrared vision array capture area is determined to obtain the infrared contour clarity distribution; S112: Based on the infrared contour clarity distribution, compare the brightness histogram distribution in the images acquired by the multi-spectral synchronous sensor array at different time points, calculate the density of texture directions in each area, divide the facial key areas by analyzing the aggregation and closure characteristics of boundary pixels, and obtain facial structure aggregation data; S113: Based on the facial structure aggregated data, feature point coordinates and symmetry mapping sets in the image frame are screened, and multi-channel fusion and superposition are performed to obtain multispectral facial features.

[0008] The present invention is improved in that the step of obtaining the spatial feature screening sequence is specifically as follows: S211: Based on the multispectral facial features, analyzing the arrangement of each feature point on the horizontal and vertical axes, determining whether its spatial distribution has the balanced and symmetrical characteristics of the facial structure, selecting image frames with consistent spatial distribution patterns, and obtaining a set of symmetrically distributed image frames; S212: Based on the symmetrically distributed image frame set, comparing the spatial orientation and connection relationship of the boundary lines in the same area, calculating the continuity of key contour lines and boundary connection between frames, and selecting frame images with consistent contour connection patterns to obtain a contour-continuous image frame set; S213: Based on the contour continuous image frame set, determine the coordinate changes of the feature points of each frame, analyze the movement trend of the key points between adjacent image frames, and select image frame sequences with stable feature point movement patterns to obtain a spatial feature screening sequence.

[0009] The present invention is improved in that the steps of obtaining the three-dimensional trajectory sequence are specifically as follows: S311: Based on the spatial feature screening sequence, extract the plane coordinates of each feature point in the X-axis and Y-axis in consecutive frames, compare the coordinate change amplitude of the same feature point in adjacent frames, identify its change trend in the time series, and establish an XY axis trend sequence; S312: Based on the XY axis trend sequence, the distribution of the Z axis coordinates of each set of feature points in the continuous frames is detected. Combined with the continuous Z axis coordinates in the time series, the mutation segments are identified, and interpolation and smoothing processing is performed to obtain the trajectory change parameters of each set of feature points in the three-axis space, and the abnormal trajectory is corrected to obtain a Z axis trajectory correction set; S313: Based on the Z-axis trajectory correction set and in combination with the trajectory continuity of the feature points in the three-axis space, the trajectory segments with balanced spatial variation amplitudes are screened, and their continuous spatial coordinates in the time series are obtained to obtain a three-dimensional trajectory sequence.

[0010] The present invention is improved in that the step of obtaining the action linkage matching data is specifically as follows: S411: Based on the three-dimensional trajectory sequence, the X-axis and Y-axis position data of the feature points of each frame are selected, and combined with the continuous time series of the Z-axis trajectory, the head rotation angle change data is extracted, and the angle difference of the facial region displacement trajectory of each frame is calculated. The posture transformation angle in each time series is paired with the corresponding feature point coordinate offset to obtain angle displacement paired data; S412: Based on the angle displacement paired data, detecting micro-motion speed information of each region, comparing the coordinate change amplitude and time interval of each region, and calculating the rate joint offset trend; S413: Based on the rate joint offset trend, perform data comparison on each group of micro-motion data and the spatial motion changes of the feature points, analyze the correlation between the structural similarity parameters and the motion linkage indicators in the combination, collect data on the combinations that meet the linkage conditions, and obtain motion linkage matching data.

[0011] The present invention is improved in that the steps further include: S5: Based on the action linkage matching data, the correspondence between the numbered feature points and the behavior parameters is analyzed, the consistency of the data structure of each group of recognition channels is determined, the joint recognition integrated port discrimination process is optimized, the group with both structure and behavior meeting the standards is identified, the identification number is marked, and the face recognition mapping result is obtained; The face recognition mapping result includes an identity mapping identifier, an identification identifier, and a feature linkage mapping index.

[0012] The present invention is improved in that the steps of obtaining the face recognition mapping result are specifically as follows: S511: Based on the action linkage matching data, analyzing the correspondence between the numbered feature points and the behavior parameters, determining the synchronization between the continuous distribution of each feature point in the three-dimensional space and the behavior change, screening corresponding groups with highly consistent structural features and behavioral features, and obtaining corresponding distribution data; S512: Based on the corresponding distribution data, compare the spatial structure relationship and behavior change pattern of the difference data group, optimize the correspondence rules between the arrangement order of the feature point combination and the behavior parameters, select the combination with consistent structure and behavior, and obtain structure mapping combination data; S513: Based on the structure mapping combination data, determine the linkage distribution between the feature point group and the behavior parameter group, calculate the participation ratio of each group of feature points and the linkage characteristics of the behavior parameters, mark the synchronization relationship and mapping index, and obtain the face recognition mapping result.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are: In the present invention, facial detail level discrimination is achieved through multispectral information fusion. Based on regional pixel distribution and structural sequence, combined with dynamic spatial trajectory analysis and abnormal motion correction, continuous spatial features and temporal behavioral parameters are matched in parallel, and behavioral linkage and facial structural features are screened synchronously. The multi-parameter joint recognition strategy in dynamic scenes reduces the interference caused by environmental changes and human intervention, improves the reliability of target identity determination in complex places, and enhances the practicality and real-time recognition capabilities for multi-source scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of the main steps of the present invention; Figure 2 This is a flowchart of obtaining multispectral facial features in the present invention; Figure 3 This is a flow chart for obtaining a spatial feature screening sequence in the present invention; Figure 4 This is a flowchart for obtaining a three-dimensional trajectory sequence in the present invention; Figure 5 This is a flowchart for obtaining action linkage matching data in the present invention; Figure 6 This is a flowchart for obtaining face recognition mapping results in the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0016] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description. They do not indicate or imply that the devices or elements referred to must have a specific direction, be constructed and operate in a specific direction, and therefore should not be understood as limiting the present invention. In addition, in the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0017] Example

[0018] See also Figure 1 The present invention provides a technical solution: an AI-based face recognition security method, comprising the following steps: S1: Based on the monitoring entrance and exit recognition channel, the pixel distribution of facial image data is analyzed to determine the clarity of the facial contour in the area captured by the infrared vision array. By comparing the image details collected by the multispectral synchronous sensor array, the facial area under the acquisition time sequence is structured segmented. Frames with complete feature point coordinates and symmetry parameters are selected to determine the distribution range of the skin color area and obtain multispectral facial features. S2: Based on multispectral facial features, the image recognition buffer unit is called to optimize the multi-frame images output by the facial feature aggregation channel, determine whether the spatial distribution of feature points in each frame is uniform, compare the structural overlap between the contour boundaries of each frame, and select facial image frames with continuous structures and spatial differences within a preset range to obtain a spatial feature screening sequence; S3: Based on the spatial feature screening sequence, the 3D spatial changes of feature points between frames are analyzed. By comparing the changing trends of the X-axis and Y-axis positions in consecutive frames, the coherence of the Z-axis trajectory in the time series is judged. Feature points with stable position changes are screened, and the trajectory of points with sudden changes in the spatial trajectory is corrected to obtain a 3D trajectory sequence. S4: Based on the 3D trajectory sequence, access the behavioral parameter comparison interface, calculate the change in head rotation angle, determine the speed of facial micro-movements recorded by the local motion capture unit, compare the change pattern of each set of motion parameters with the spatial coordinates of the feature points, and select data that meets the matching requirements in terms of structural similarity and motion linkage, to obtain motion linkage matching data; S5: Based on the action linkage matching data, analyze the correspondence between the numbered feature points and the behavioral parameters, determine the structural consistency of each group of data in the identity recognition channel, optimize the discrimination process of the joint recognition integrated port, identify the feature groups whose structure and behavior meet the standard requirements, compare the distribution ratio of the linkage features of each group, and mark each group of identification numbers to obtain the face recognition mapping results.

[0019] Multispectral facial features include spectral distribution parameters, texture features, and skin color block information; spatial feature screening sequences include spatial consistency indicators, structural aggregation labels, and inter-frame stable segments; three-dimensional trajectory sequences include coordinate trajectory data, continuity markers, and corrected trajectory sets; action linkage matching data includes behavior linkage parameters, action recognition numbers, and linkage matching relationship groups; and face recognition mapping results include identity mapping identifiers, recognition identifiers, and feature linkage mapping indexes.

[0020] In S1, the identification channel refers to a physical channel or virtual detection area with face collection and recognition functions deployed in security scenarios, which is commonly found in access control, entrances and exits, checkpoints and other monitoring deployment sites; the infrared vision array refers to an array device composed of multiple infrared sensor units, which is used to capture infrared images of faces in low light, backlight or nighttime environments to enhance the adaptability of face detection; the multi-spectral synchronous sensor array refers to an integrated sensor module that can simultaneously collect multiple spectral images such as visible light, near infrared, and short-wave infrared, ensuring the acquisition of multi-source face image information under different lighting and camouflage conditions; structure Digital segmentation refers to the precise separation of the facial area from the panoramic or background image, and further divides it into specific functional areas (such as eyes, nose, mouth, face shape, etc.) for subsequent precise feature analysis; a complete frame refers to an image frame in which the information of key facial features (such as facial features and contours) is not blocked or missing, and the image quality meets the analysis requirements, which is suitable for subsequent feature extraction and comparison; the skin color area distribution range refers to the skin tone and skin color distribution area boundaries determined by pixel color analysis within the facial area, which helps to assist in the investigation of occlusion or camouflage and improve recognition robustness.

[0021] In S2, the image recognition buffer unit refers to a buffer or module used to temporarily store and manage multi-frame facial image data, which facilitates continuous processing and screening and improves the real-time performance and stability of the system; the facial feature aggregation channel refers to a dedicated processing channel that aggregates, fuses and aligns facial features (such as key points, area segmentation results, etc.) collected from multiple frames; whether the distribution is uniform is to judge whether the distribution of key facial feature points (such as eyes, nose, mouth, etc.) in the image space is regular and has no abnormal offset, so as to avoid misjudgment caused by abnormal detection; structural overlap refers to comparing the degree of overlap of the boundaries of key facial parts in different frame images, measuring the consistency of multiple frame images, and eliminating recognition errors caused by jitter, blur and other factors; spatial difference is in a preset range refers to judging whether the amplitude of the change in the spatial position of the same feature point in multiple frame images is within the normal range defined by the system, ensuring that the screened frames are highly similar.

[0022] In S3, three-dimensional spatial change refers to the position change process of facial feature points in the three spatial dimensions of X, Y, and Z (plane position + depth), reflecting changes in facial movements or postures; change trend refers to analyzing the spatial coordinate change trajectory of feature points in multiple consecutive frames to determine whether it presents specific trends such as stability, linearity, periodicity, or abnormal offset; coherence refers to measuring the continuity and stability of the three-dimensional trajectory in the frame sequence to avoid abnormal phenomena such as mutations and jumps, and to ensure the accuracy of feature tracking; stable feature points refer to facial feature points with small spatial position changes, smooth trajectories, and not easily disturbed in multiple frames of data, which are suitable for use as a basis for identity discrimination; spatial trajectory refers to the motion path formed by feature points over time in a three-dimensional coordinate system, which is used for dynamic analysis and identity recognition; trajectory correction refers to the adjustment of feature point trajectories that have abnormal fluctuations or mutations by algorithm smoothing, interpolation, or outlier removal to ensure the consistency of the overall trajectory.

[0023] In S4, the behavioral parameter comparison interface refers to the data interface or processing unit in the system used to access and process parameters related to facial movements (such as head rotation, expression changes, etc.); the local motion capture unit is an acquisition module specifically used to collect and record changes in facial micro-movements (such as twitching of the mouth corners, raising of eyebrows, etc.) at high frequency, thereby improving the dynamic behavior analysis capability; the speed of facial micro-movements refers to the speed characteristics of the movement of local facial parts (such as eyes, mouth, eyebrows, etc.) measured through time series analysis, which is an important basis for distinguishing between living and disguised objects; the change pattern refers to the description of the change law between the action parameters and the spatial coordinates of the feature points over time, such as synchronization, delay, mutation and other patterns; structural similarity refers to the degree of similarity of key structural features between quantitatively compared data samples, and is often used to screen action data of the same identity or high correlation; motion linkage refers to the analysis of whether there is a coordinated change relationship between the head and local facial movements, which is used to determine the naturalness of the movement or whether there is any forgery; the matching requirement refers to the minimum standard for the degree of correlation between structural features and behavioral features during the recognition process, and data below this standard will be excluded.

[0024] In S5, numbered feature points refer to the assignment of unique numbers to the key points collected from each facial image, which facilitates the accurate tracking and identification of corresponding points during multi-frame and multi-parameter fusion; corresponding relationships refer to the one-to-one correspondence and mutual correlation between feature point data and behavioral parameter data, which facilitates joint analysis and multi-parameter comprehensive judgment; identity recognition channel refers to the data processing channel used to make final identity recognition decisions on candidate facial data, which is the core judgment link; structural consistency refers to the degree of consistency of the same candidate identity in various feature dimensions under multi-parameter fusion. High structural consistency indicates high credibility of identity judgment; joint recognition integration port refers to the interface or data endpoint that summarizes multi-source data such as multi-dimensional recognition results, behavioral actions, spatial features, etc. for comprehensive output and judgment; feature groups that meet standard requirements refer to identity feature groups that have high credibility and uniqueness through the combination of feature points that pass all structural and behavioral screening criteria; marking each group of identification numbers refers to assigning a unique identification number to each identity candidate group for subsequent output and management.

[0025] See also Figure 2 ,The steps for acquiring multi-spectral facial features are as follows: S111: Based on the monitoring entrance and exit recognition channel, the pixel distribution of the facial image data is analyzed. For each block of the image sequence, the distribution density of the edge continuity of each pixel block is calculated by statistically analyzing the change level of the grayscale gradient. The edge clarity change trend of each area in the infrared vision array capture area is determined to obtain the infrared contour clarity distribution; Perform pixel-level analysis on the facial image sequence collected by the infrared vision array, divide each frame of the image into multiple pixel block areas according to a fixed grid, and each pixel block can be set to 32 by 32 pixels. The grayscale value change information in each pixel block is extracted block by block, and the grayscale value difference sequence between each group of adjacent pixels is counted. By comparing the change amplitude of the difference, it is determined whether the current pixel block has obvious edge features. Further, in each pixel block, the number of pixel pairs whose grayscale difference change amplitude reaches or exceeds the set reference value is counted, and the proportion of this number in the entire pixel block is calculated. If the proportion exceeds 30%, it is determined to be an area with high edge continuity. Indexes are established for all pixel blocks that meet this condition, and their corresponding spatial positions in the entire image are recorded. Then, by traversing the image space, the number of pixel blocks that meet the edge continuity condition in the unit area is counted. , forming the spatial distribution data of edge clarity, judging whether the area is an edge clear area according to the standard of whether the number of clear pixel blocks in each area reaches 600, extracting all edge clear areas in the image and merging their contours, analyzing their arrangement and distribution density changes in the image space, recording the spatial boundary positions of each area, forming an edge distribution trend map, and processing the continuous frames of the image sequence at the same time, comparing the changes in the number of edge clear areas in the frames collected at different time points at the same spatial position, and counting whether the change rate is stable. If the clarity change ratio of a region in all frames is less than 5%, and the area ratio it occupies in the image exceeds 70%, then the region is judged to be a stable clear area under infrared capture, and the distribution trend data of the edge contour clear area of ​​the entire image under the infrared vision array is obtained as the infrared contour clarity distribution amount.

[0026] S112: Based on the infrared contour clarity distribution, compare the brightness histogram distribution in the images acquired by the multispectral synchronous sensor array at different time points, calculate the density of texture directions in each area, and divide the facial key areas by analyzing the aggregation and closure characteristics of boundary pixels to obtain facial structure aggregation data; The image brightness comparison analysis is performed on the face images collected by the multispectral synchronous sensor array at different times. The images taken at two time points are selected and grayscale processed. The brightness information in the image is extracted as a distribution matrix. The image brightness distribution is then normalized to compare the illumination changes between the images. The image brightness distribution at the two time points is then plotted as a histogram. The difference in the frequency of each brightness level in the two frames of images is calculated. The difference of all brightness levels is averaged as a comparison indicator of the overall brightness change. If the average difference exceeds 15, it is considered that there is a significant brightness difference between the images. At the same time, the area corresponding to the brightness difference is marked in the image space. After that, the image is edge-corrected. Extraction processing, statistics of texture direction information in each area, by traversing the grayscale change direction of each pixel in its adjacent area, determine whether the dominant texture direction in the area is concentrated, if the proportion of the dominant direction exceeds 80%, the area is judged as a texture direction dense area, then judge the aggregation state of the boundary pixels, and judge its closed characteristics by detecting whether the edge pixels form a closed contour. If the proportion of closed boundary pixels is greater than 60%, the area is marked as a boundary closed area. Finally, the areas in the image space that meet the requirements of significant brightness changes, concentrated texture directions, and closed boundary structures are collectively processed, and their spatial position and structural feature data are output to obtain facial structure aggregation data for subsequent structural analysis.

[0027] S113: Based on the facial structure aggregation data, the feature point coordinates and symmetry mapping set in the image frame are screened using the formula: ; Perform multi-channel fusion and superposition to obtain multispectral facial features ,in, Indicates the The symmetry combination factor of the feature points, Indicates the The brightness frequency band response of the feature points in the left area after structure division, Indicates the The brightness frequency band response of the feature points in the right area after structure division, Indicates the The degree of texture interference within the region, represents the spectrum collaborative fusion enhancement term, Represents the superposition of visible light spectrum reflectance measurement, represents the absorption response fusion amount in the near-infrared band, Indicates skin color area coverage mapping adjustment item, Indicates the pixel coverage composite area of ​​the skin color block area, Indicates the number of feature points.

[0028] Multispectral facial features are a fusion of multiple spectral channel information such as visible light and near-infrared in the same facial area, combined with multiple parameters such as feature point symmetry, brightness frequency band response of left and right areas, texture interference parameters of each local area, and pixel coverage area of ​​skin color area. After fusion operation, the comprehensive expression of facial image features obtained is used for identity feature mapping, feature point screening, structural consistency judgment and other processing processes in subsequent face recognition systems. It is a multi-dimensional basic data indicator for distinguishing individual identity and detecting camouflage or occlusion.

[0029] Extract feature points from each key area marked in the image frame sequence and divide them into numbered sets. For each numbered feature point, extract the brightness response of the left and right symmetrical areas of the face under the visible light spectrum channel. 、 , measure the texture interference degree of the area where the point is located , and give it a symmetry combination factor , perform multi-channel fusion superposition on each parameter according to the following formula, and set the number as The three feature points have the following original values: 、 、 、 ; 、 、 、 ; 、 、 、 ; Set normalized values ​​for visible light reflectance measurement superposition , the near-infrared absorption response fusion amount is normalized to , the fusion enhancement item is set to , Skin color area mapping adjustment , the skin color block pixel coverage area is normalized and the value is .

[0030] The calculation is as follows: Item 1: ; Item 2: ; Item 3: ; Sum the above three items and we get the weighted structure item: ; Then calculate the spectral enhancement term: ; The numerator is: ; The denominator is: ; Multispectral facial features are: ; The results show that the currently calculated multispectral facial feature image quantity GS≈6.11 is higher than the set standard benchmark interval upper limit of 4.5. This benchmark interval is obtained through the statistics of the fusion feature expression results of large-scale facial image samples, and is usually distributed in the range of 2.0~4.5. Among them, GS <2.0 is marked as insufficient feature expression, and GS >4.5 indicates high integrity of the fusion feature, stable structural matching, and active spectral response. This numerical result shows that the feature points extracted from the current image frame show a high degree of synergy in the responses of structural symmetry, regional texture and multispectral channels, so the overall fusion value is significantly amplified, indicating that it has the basic judgment qualification to be directly included in the next stage of spatial screening and recognition comparison process. Subsequently, this feature quantity can be used to participate in the spatial trajectory stability analysis and dynamic behavior comparison process.

[0031] See also Figure 3 ,The steps for obtaining the spatial feature screening sequence are as follows: S211: Based on the multispectral facial features, analyze the arrangement of each feature point on the horizontal and vertical axes to determine whether its spatial distribution has the balanced and symmetrical characteristics of the facial structure, select image frames with consistent spatial distribution patterns, and obtain a set of symmetrically distributed image frames; Based on multispectral facial features, the facial feature data includes the horizontal and vertical coordinate positions of the feature points extracted from each frame image, the distribution of skin color areas, the positions of texture lines, etc., extract no less than 50 feature points from the face image, and record their X-axis and Y-axis coordinate values ​​in the image's two-dimensional coordinate system respectively. First, a coordinate table is independently established for each frame image, and the coordinate extremes of all feature points in the horizontal axis direction are calculated. With the facial central axis (such as the vertical direction where the center point between the two eyes is located) as the reference line, it is determined whether the number of feature points on the left and right sides is equal, and whether the horizontal axis coordinates of the corresponding points on both sides are distributed in pairs about the central axis, that is, whether the symmetric difference is less than the set symmetric error threshold. For example, the threshold is set to 3 pixel points, the symmetrical difference is the difference between the two corresponding points and the center line. If the proportion of all paired feature points that meet the symmetrical difference less than the threshold reaches more than 80%, the image frame is determined to have symmetrical characteristics. The distribution of feature points in the longitudinal direction is further evaluated, and it is recorded whether each point is equally spaced or evenly arranged in the vertical direction. It is determined whether there is a concentrated or abnormally concentrated area. If the horizontal and vertical coordinate distributions meet the above structural balance and symmetry requirements at the same time, the frame image is added to the preliminary set, and each frame in the image sequence is traversed. The above operation is repeated to screen out image frame samples that meet the spatial symmetrical structural characteristics, and record them as a symmetrically distributed image frame set by establishing an index.

[0032] S212: Based on the symmetrically distributed image frame set, the spatial orientation and connection relationship of the boundary lines in the same area are compared, the continuity of the key contour lines and the boundary connection between the frames are calculated, and the frame images with the same contour connection pattern are selected to obtain the contour continuous image frame set; The spatial structure of the key facial boundary lines in each frame image is evaluated in turn. The key boundary lines refer to the nose bridge contour line, facial side line, mandibular contour line, etc. extracted from the image. First, the adjacent connection position of each pixel on the boundary line is extracted according to the pixel coordinates to generate a continuous direction sequence of boundary pixels. Then, the extension direction difference of the boundary lines in different frame images of the same facial area is compared, and the coordinate difference is used to determine the directional consistency of the boundary direction. For example, the direction vector difference of the nose bridge contour line extending downward in frame 1 and frame 2 is taken. If the angle between the two vectors is less than 10 degrees, it is determined that the direction is consistent. If the angle is greater than 15 degrees, it is determined that the difference is obvious. Then the boundary line connection is calculated. The density of connection points is used to determine whether the contour line is broken, and the ratio of the number of connection points to the line segment length is taken as a continuity reference indicator. If the ratio is higher than the set threshold (for example, there are more than 8 continuous connection points per 10 pixels), the boundary connection is considered to be good. The coordinates of the start and end points of the boundary line in each frame are further compared. If the offset distance between the start and end points is less than 5 pixels between the two frames, the boundary alignment is determined to be good. By cross-judging the consistency of the direction vector, connection density and start and end coordinates of the same boundary line in all image frames, the image frames that meet the requirements of the above three judgment conditions are screened out and recorded as frames with consistent contour connection mode, which are collected to form a contour continuous image frame set.

[0033] S213: Based on the continuous contour image frame set, determine the coordinate changes of the feature points of each frame, analyze the movement trend of the key points between adjacent image frames, and select an image frame sequence in which the motion pattern of the feature points is stable to obtain a spatial feature screening sequence; The position change of feature points between image frames is analyzed, and the coordinates of all key feature points in each frame are extracted in turn. The corresponding relationship between the points and the previous frame is established to determine whether there is confusion or omission of corresponding point numbers. First, the difference of the coordinate change value of each feature point in the X-axis direction in consecutive frames is calculated, and then the same process is performed on the Y-axis direction to obtain the displacement vector sequence of each feature point. By counting the average displacement distance of all feature points, it is recorded whether it is in a stable change range. The range threshold is set to a maximum movement of no more than 5 pixels per frame. If the displacement distance of a feature point is less than this value in three consecutive frames, its motion trend is determined to be stable. Then, the number of points that meet the stability condition among all feature points is counted. If the proportion of their number to the total number of feature points exceeds 85%, the frame image is recorded as a stable frame of feature point motion law. All frames are traversed and the above judgment is repeated. By establishing a sequence index set of stable frames, image frames with high feature point position stability are arranged and combined to form a spatial feature screening sequence.

[0034] See also Figure 4 , the specific steps for obtaining the three-dimensional trajectory sequence are: S311: Based on the spatial feature screening sequence, extract the plane coordinates of each feature point in the X-axis and Y-axis in consecutive frames, compare the coordinate change amplitude of the same feature point in adjacent frames, identify its change trend in the time series, and establish an XY axis trend sequence; Each frame image in the sequence has passed the feature structure screening and has the characteristics of continuous and stable distribution of facial feature points. The X-axis and Y-axis values ​​of the key feature points in the plane coordinate system are extracted in turn for all the frame images. The coordinates of each feature point in the corresponding image are expressed as two-dimensional pixel positions. For example, a point is (120, 250) in frame 1 and (122, 252) in frame 2. A one-to-one correspondence between the feature point numbers between frames is established in chronological order. The X-axis and Y-axis coordinates of the same numbered point in each pair of adjacent frames are calculated separately, and the movement amplitude in both directions is recorded. If the X-axis change amplitude of a point is 2 pixels and the Y-axis is 2 pixels, then the point is offset in the plane. This process is executed one by one for all feature points and a complete difference matrix is ​​generated. Then, the number of feature points in each consecutive frame is counted. The trend of change is determined by judging whether its change direction (for example, the X coordinates of three consecutive frames are all shifted to the right) and amplitude change are within the stable range. The stable range is set as the coordinate difference between each frame is within 3 pixels, and the change direction is continuous and does not reverse. If the proportion of feature points that meet this condition exceeds 80% within 10 frames, the trend of the point is considered stable. All feature points with stable trends are further numbered and indexed. The trajectory curve of each point is drawn according to its plane coordinate change over time. The change direction, fluctuation amplitude and continuity mark information of each curve are extracted. The abnormal change segments are marked and excluded. The remaining continuous change parts are numbered and saved to the trend set. The trend information of all feature points is organized in frames to form an XY axis trend sequence that reflects the evolution characteristics of key points in the plane direction over time.

[0035] S312: Based on the XY axis trend sequence, the distribution of the Z axis coordinates of each set of feature points in the continuous frames is detected. Combined with the continuous Z axis coordinates in the time series, the mutation fragments are identified and interpolation smoothing is performed. The formula is used: ; Obtain the trajectory change parameters of each set of feature points in the three-axis space, and correct the abnormal trajectory to obtain the Z-axis trajectory correction set, where: Indicates the The trajectory change parameters of the group feature points, Indicates the The feature points of the group The Z-axis coordinate of the frame, Indicates the The feature points of the group The Z-axis coordinate of the frame, Indicates the The feature points of the group The X-axis coordinate of the frame, Indicates the The feature points of the group The X-axis coordinate of the frame, Indicates the The feature points of the group The Y-axis coordinate of the frame, Indicates the The feature points of the group The Y-axis coordinate of the frame, Indicates the total number of frames in the time series; The trajectory change parameter refers to a comprehensive parameter obtained by measuring the spatial trajectory fluctuation and change amplitude of the same set of feature points in three-dimensional space (X, Y, and Z axes) over time in consecutive frames. It reflects the stability and degree of change of the spatial trajectory of the feature points in the entire time series. This parameter is used to identify and correct abnormal trajectory segments.

[0036] Detect the distribution of the Z-axis coordinates of each set of feature points in consecutive frames, obtain the three-dimensional spatial coordinates of each set of feature points in the five-frame sequence, and construct the corresponding spatial trajectory sequence ,in 、 Derived from the positioning coordinates of the feature points of the two-dimensional image, The pixel depth map data obtained by the multispectral depth sensor is extracted and normalized to participate in subsequent operations. The Z-axis depth change and X-axis lateral displacement between adjacent frames are calculated in sequence, and the inter-frame movement amplitude of the Y axis is used as the combined denominator. The feature point number is set as , the sampling results in the 5-frame time series are as follows: The original value of the Z axis is: [138, 142, 146, 150, 149]; The X axis is: [62, 64, 66, 67, 66]; The Y axis is: [48, 47, 45, 44, 43]; The corresponding data after normalization is: : [1.380, 1.420, 1.460, 1.500, 1.490]; : [0.620, 0.640, 0.660, 0.670, 0.660]; : [0.480, 0.470, 0.450, 0.440, 0.430]; Substitute the formula and calculate item by item as follows: ; ; ; ; Calculate the feature points The trajectory change parameters within the 5-frame time period are If the preset upper limit of the trajectory correction range is 0.05, the parameter is within the correctable area, indicating that the spatial variation of the feature point during the observation period is acceptable. There is a certain degree of jump, but it is not enough to constitute a trajectory break. Interpolation correction can be performed, and then the local fluctuation area is smoothed to generate a continuous Z-axis trajectory correction set for subsequent 3D trajectory sequence construction steps.

[0037] S313: Based on the Z-axis trajectory correction set and combined with the trajectory continuity of the feature points in the three-axis space, the trajectory segments with balanced spatial variation amplitudes are screened, and their continuous spatial coordinates in the time series are obtained to obtain a three-dimensional trajectory sequence; Each feature point trajectory has been subjected to abnormal change point removal and trajectory reconstruction operations. The time series trajectory data of the feature points in the three-dimensional coordinate system is read in sequence, and each trajectory line is analyzed in the order of the feature point number. The X, Y, and Z coordinates recorded in each frame are traversed horizontally, and the difference in the change amplitude in the three directions between consecutive frames is recorded. Then, whether the three-axis change amplitude of the point in the continuous time is determined to be stable. The judgment standard is set as follows: if the coordinate change difference between each frame and the previous frame in each axis does not exceed 4 pixels in 5 consecutive frames of data, the trajectory change amplitude is determined to be balanced and marked as a candidate trajectory segment. All candidate trajectory segments are numbered and summarized. If a single segment is at least 7 frames long and the three-axis coordinate change ratio does not deviate from the benchmark ratio by ±20%, the segment is included in the valid trajectory sequence. Further, the timeline of all valid segments is reconstructed and they are arranged continuously with the frame number as the index. The coordinate points in the X, Y, and Z directions of each frame in the sequence are extracted, and a point array structure in 3D space is constructed. The structure is saved as a 3D trajectory sequence, which serves as the basic data for subsequent behavior matching and action recognition.

[0038] See also Figure 5 , the specific steps for obtaining action linkage matching data are: S411: Based on the three-dimensional trajectory sequence, the X-axis and Y-axis position data of the feature points of each frame are selected, and combined with the continuous time series of the Z-axis trajectory, the head rotation angle change data is extracted. The angle difference of the facial region displacement trajectory of each frame is calculated, and a pairing relationship is established between the posture transformation angle in each time series and the corresponding feature point coordinate offset to obtain angle displacement pairing data; Each feature point has X-axis, Y-axis and Z-axis coordinate values ​​in continuous frames. For all feature points in each frame image, first extract the plane position of the X-axis and Y-axis. By setting the facial center point as the reference point, standardize the position offset direction and amplitude of each feature point. Then, combine the curve of the change of the Z-axis coordinate over time to judge the rotation relationship between the displacement path of each feature point in three-dimensional space and the overall posture of the head. Set the reference axis of the head posture change as the Z-axis vertical line. Take 3 frames as the time window, extract the feature point coordinates of the starting frame, middle frame and end frame respectively, and calculate the rotation angle change of the line connecting any two feature points in the three frames by the vector angle method. The angle change values ​​obtained from all feature points are averaged to form the overall posture transformation angle of the frame. It is judged whether it rises or falls continuously in the time series. If there are three consecutive frames with the same angle change trend, the trend change is recorded as a valid rotation sequence. The rotation angle change data is then paired with the spatial coordinate difference of the corresponding feature point. The posture change angle value in each frame is combined with the offset of the key feature point in the frame one-to-one. The corresponding frame number, angle value, corresponding point number and its three-dimensional offset value are recorded to form a pairing data set containing the angle displacement mapping relationship. The time axis data structure is constructed with the frame number as the main index to obtain the angle displacement pairing data.

[0039] S412: Based on the angle displacement pairing data, detect the micro-motion speed information of each area, compare the coordinate change amplitude and time interval of each area, and use the formula: ; Calculate rate joint migration trend ,in, Indicates the total number of regions involved in the comparison, Indicates the Micro-motion speed data obtained from area detection, Indicates the The coordinate change range of the regional feature points, Indicates the The time interval between adjacent frames in a region; The rate joint deviation trend is the overall deviation between the micro-motion speed of facial feature points and the spatial motion speed in all comparison areas, which is used to reflect the consistency or coordination between multi-region motion and coordinate changes.

[0040] For the five facial regions defined in the surveillance scene, the micro-motion speed information recorded by the local motion capture unit in each region is detected one by one. The spatial coordinate change amplitude and time interval of each key feature point are calculated by combining the previous and next frame image data. The micro-motion speed is compared with the unit time coordinate displacement rate of the region. The rate combined with the offset trend formula is used to calculate the offset of the above data, where: , indicating that there are five comparison areas. Indicates the The micro-movement speed of the area, in mm / s, Indicates the coordinate change range of the area between two consecutive frames, in mm. It represents the time interval between frames, in seconds. Now assume that the following raw data are obtained for regions 1 to 5 respectively: 、 、 、 、 ; The corresponding displacement amplitude and time interval are: 、 、 、 、 (All units are in mm); 、 、 、 、 (All units are in seconds); After normalizing the units of each region, the following rate terms are obtained: The first area is , ; The second area is , ; The third area is , ; The fourth area is , ; The 5th area is , ; After substituting them into the formula, we can get: ; ; This result shows that the current calculated rate joint deviation trend Compared with the set motion linkage offset judgment benchmark interval [0, 1.2], it is within the allowable matching tolerance range, indicating that the overall difference between the micro-motion speed and the feature point displacement rate among the five regions does not exceed the upper limit of the judgment standard, and is in a state of strong linkage response consistency, which means that all current regions have a strong coordination basis in structural motion synchronization, which can be used as a reference for the next step of motion and structure matching judgment, and used to assist in establishing a valid combination sequence in the motion linkage matching data.

[0041] S413: Based on the rate joint offset trend, perform data comparison on each set of micro-motion data and the spatial motion change of the feature points, analyze the correlation between the structural similarity parameters and the motion linkage index in the combination, collect data on the combinations that meet the linkage conditions, and obtain motion linkage matching data; Obtain the corresponding timestamps of each set of motion data captured by the facial micro-motion recording unit and the spatial change data of the feature points, align the micro-motion timing with the three-dimensional trajectory data through synchronization processing, first extract the key action nodes associated with each set of micro-motion events, such as the corner of the mouth rising, the eyebrow moving down, etc., and extract the spatial coordinate offset value and movement direction of the feature points in the three frames before and after the action occurs, record the offset rate in the X, Y, and Z directions respectively, bind the spatial feature point number corresponding to each set of micro-motion data, and count whether it is consistent with the trend direction in the feature point trajectory at the same time sequence. If the feature point coordinate offset direction corresponding to the micro-motion occurrence frame is consistent with the trend of the previous and next frames, and the change rate is consistent with the trend of the previous and next frames, the feature point number will be used as the tracking number. If the rate is within the range of ±15% of the set threshold, the group is recorded as a linkage sample. Then, a structural similarity analysis is performed on the distribution morphology of the feature points in all samples, and the difference between the relative distance relationship between the feature points and the standard facial structure model is counted. If the average value of the difference is less than the set structural similarity threshold of 3 pixels, and the area where the micro-motion occurs overlaps with the linkage feature point distribution area for more than 80%, the data combination is considered to meet the linkage conditions. Finally, all sample combinations that meet the above-mentioned offset rate matching, direction consistency, and structural similarity are extracted to generate four types of information items: action number, timestamp, linkage point number set, and spatial offset value, which are organized into action linkage matching data.

[0042] See also Figure 6 ,The steps to obtain the face recognition mapping results are as follows: S511: Based on the action linkage matching data, the correspondence between the numbered feature points and the behavior parameters is analyzed, the continuous distribution of each feature point in the three-dimensional space and the synchronization of the behavior changes are determined, and the corresponding groups with highly consistent structural features and behavioral features are selected to obtain corresponding distribution data; The data contains each set of feature point numbers and the corresponding behavioral parameter sequence. First, the continuous frame coordinate trajectories of all feature points in three-dimensional space are extracted, and a time series index table is established for the X, Y, and Z coordinates of each point. The trajectory segments of each set of feature points within 10 consecutive frames are extracted, and the position change vectors between frames are recorded. Then, the behavioral parameter data under the same time series are synchronously positioned, and the change marks of action parameters such as "eyebrow raising", "mouth corner deviation", and "head turning left" are extracted. It is analyzed whether the trajectory change direction of each feature point is synchronized with the time series change of the behavioral parameter. For example, from the 5th to the 7th frame, if the Y-axis coordinate of the feature point number 23 continues to rise and the amplitude exceeds 3 pixels, and the marked action during this period is If the eyebrow is raised, it is determined that number 23 is synchronized with the behavior parameter. The above judgment is performed on all numbered feature points, the numbers for which the synchronization judgment is established are marked, and the pairing relationship between them and the action label is recorded. The synchronized pairs are further screened for structural consistency. By counting the change difference of their three-axis motion trajectory in all valid frames to see whether it is controlled within the threshold range, if the motion direction of three consecutive frames remains consistent and the offset value difference of each frame does not exceed 4 pixels, the group of structural changes is determined to be stable. The numbered behavior pair combination with this feature is retained in the screening result set. Finally, a mapping table with one-to-one correspondence between numbered feature points and behavior parameters is constructed to obtain corresponding groups with highly consistent structures and behavior parameters, forming corresponding distribution data.

[0043] S512: Based on the corresponding distribution data, compare the spatial structure relationship and behavior change pattern of the difference data group, optimize the corresponding rules between the arrangement order of the feature point combination and the behavior parameters, select the combination with consistent structure and behavior, and obtain structure mapping combination data; The selected feature points and their corresponding behavior parameters are further divided into different sub-combinations. By comparing the relative coordinate structure of the feature points in the three-dimensional space in each combination, it is determined whether there is a shape offset in the spatial arrangement between the point groups. The mean relative distances of the feature points with the same number in the two groups in the three-axis space are calculated respectively, and then the difference is compared to see whether it exceeds the set structural similarity threshold of 5 pixels. For example, the mean value of the feature point group A in the Z-axis direction is 24, and that of group B is 29, then the structural offset is 5. If the difference is greater than the threshold, it is marked as a difference combination. Then, the change pattern of the behavior parameter corresponding to the difference combination is analyzed to determine whether the trend of the parameter in the frame sequence is consistent with the original combination. For example, If the behavioral parameters in group A are continuously rising, while those in group B are intermittently changing, it is recorded as inconsistent behavioral patterns. Finally, all combinations of structural differences and behavioral differences are classified, and combinations with structural deviations less than 5 pixels and consistent behavioral change trends are marked as preferred combinations. The arrangement order of feature points in this type of combination is reconstructed, sorted by the priority of three-axis coordinate stability, the numbering order is adjusted, and the new arrangement rules are recorded. At the same time, the behavioral parameters are regrouped according to their correlation with the change trend of the rearranged feature points. Pairing relationships with correlations greater than 0.85 are retained, and the rest are eliminated. Finally, the feature point combination and behavioral parameter combination pairs with consistent structure and behavior are output to generate structural mapping combination data.

[0044] S513: Based on the structure mapping combined data, determine the linkage distribution between the feature point group and the behavior parameter group, calculate the participation ratio of each feature point group and the linkage characteristics of the behavior parameters, mark the synchronization relationship and mapping index, and obtain the face recognition mapping result; The linkage relationship between each set of feature point numbers and the corresponding behavior parameters is judged. First, the ratio of the number of numbers in each set of feature points to the total number of numbers is extracted, and this ratio is set as the participation ratio. For example, the total number of feature points is 40, and the current combination involves 28 numbers, then the participation ratio is 70%. For each set of behavior parameter sequences, its change frequency and amplitude in the action trigger frame segment are extracted to determine whether it and the above feature points have a common spatial offset in the same frame. If more than 80% of the points in the same frame move in one direction at the same time, and the frame is marked as a specific action change frame, it is marked as a strong linkage relationship. Then, all feature point combinations are counted, and each set is Calculate the frequency of occurrence, behavior consistency rate and participation ratio, set the participation ratio higher than 60% and the behavior consistency rate higher than 85% as the screening criteria, and mark all combinations that meet the conditions as synchronous linkage relationship groups. Then generate the mapping index number according to the feature point number set and the behavior parameter sequence. For example, if the feature point group T17-32 is linked with the action groups A2 and A5, the index number is marked as IDX_T17-32_A2A5. The index number is output as the final identification label to form a number correspondence table and an action parameter association table, and the synchronization label status is marked to generate a face recognition mapping result with a unique identification identifier.

[0045] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. An AI-based face recognition security method, characterized in that: The following steps are involved: S1: Based on the monitoring entrance and exit recognition channel, analyze the pixel distribution of the facial image, determine the contour clarity captured by the infrared vision array, compare the acquisition details of the multispectral sensor, screen the feature points and symmetry parameters of the complete frame, determine the skin color area distribution, and obtain the multispectral facial features; S2: Based on the multispectral facial features, calling the image recognition buffer unit, optimizing the multi-frame images of the facial feature aggregation channel, judging the uniformity of the spatial distribution of feature points, comparing the overlap of contour structures, screening consecutive facial frames with spatial differences within an interval, and obtaining a spatial feature screening sequence; S3: Based on the spatial feature screening sequence, analyze the three-dimensional spatial changes of feature points, compare the X-axis and Y-axis position trends of consecutive frames, determine the Z-axis trajectory continuity, screen feature points with stable position changes, perform corrections on trajectory mutation points, and obtain a three-dimensional trajectory sequence; S4: Based on the three-dimensional trajectory sequence, calculate the change in head rotation angle, determine the speed of facial micro-movements recorded by the local motion capture unit, compare the movement parameters with the spatial coordinate change pattern of the feature points, screen the data with the best matching degree, and obtain the movement linkage matching data.

2. The AI-based face recognition security method according to claim 1, characterized in that: The multispectral facial features include spectral distribution parameters, texture features, and skin color block information; the spatial feature screening sequence includes spatial consistency indicators, structural aggregation labels, and inter-frame stable segments; the three-dimensional trajectory sequence includes coordinate trajectory data, coherence marks, and corrected trajectory sets; and the action linkage matching data includes behavior linkage parameters, action identification numbers, and linkage matching relationship groups.

3. The AI-based face recognition security method according to claim 1, characterized in that: The steps of acquiring the multispectral facial features are specifically as follows: S111: Based on the monitoring entrance and exit recognition channel, the pixel distribution of the facial image data is analyzed. For each block of the image sequence, the distribution density of the edge continuity of each pixel block is calculated by statistically analyzing the change level of the grayscale gradient. The edge clarity change trend of each area in the infrared vision array capture area is determined to obtain the infrared contour clarity distribution; S112: Based on the infrared contour clarity distribution, compare the brightness histogram distribution in the images acquired by the multi-spectral synchronous sensor array at different time points, calculate the density of texture directions in each area, divide the facial key areas by analyzing the aggregation and closure characteristics of boundary pixels, and obtain facial structure aggregation data; S113: Based on the facial structure aggregated data, feature point coordinates and symmetry mapping sets in the image frame are screened, and multi-channel fusion and superposition are performed to obtain multispectral facial features.

4. The AI-based face recognition security method according to claim 1, characterized in that: The steps for obtaining the spatial feature screening sequence are specifically as follows: S211: Based on the multispectral facial features, analyzing the arrangement of each feature point on the horizontal and vertical axes, determining whether its spatial distribution has the balanced and symmetrical characteristics of the facial structure, selecting image frames with consistent spatial distribution patterns, and obtaining a set of symmetrically distributed image frames; S212: Based on the symmetrically distributed image frame set, comparing the spatial orientation and connection relationship of the boundary lines in the same area, calculating the continuity of key contour lines and boundary connection between frames, and selecting frame images with consistent contour connection patterns to obtain a contour-continuous image frame set; S213: Based on the contour continuous image frame set, determine the coordinate changes of the feature points of each frame, analyze the movement trend of the key points between adjacent image frames, and select image frame sequences with stable feature point movement patterns to obtain a spatial feature screening sequence.

5. The AI-based face recognition security method according to claim 1, characterized in that: The steps for obtaining the three-dimensional trajectory sequence are specifically as follows: S311: Based on the spatial feature screening sequence, extract the plane coordinates of each feature point in the X-axis and Y-axis in consecutive frames, compare the coordinate change amplitude of the same feature point in adjacent frames, identify its change trend in the time series, and establish an XY axis trend sequence; S312: Based on the XY axis trend sequence, the distribution of the Z axis coordinates of each set of feature points in the continuous frames is detected. Combined with the continuous Z axis coordinates in the time series, the mutation segments are identified, and interpolation and smoothing processing is performed to obtain the trajectory change parameters of each set of feature points in the three-axis space, and the abnormal trajectory is corrected to obtain a Z axis trajectory correction set; S313: Based on the Z-axis trajectory correction set and in combination with the trajectory continuity of the feature points in the three-axis space, the trajectory segments with balanced spatial variation amplitudes are screened, and their continuous spatial coordinates in the time series are obtained to obtain a three-dimensional trajectory sequence.

6. The AI-based face recognition security method according to claim 1, characterized in that: The steps for obtaining the action linkage matching data are specifically as follows: S411: Based on the three-dimensional trajectory sequence, the X-axis and Y-axis position data of the feature points of each frame are selected, and combined with the continuous time series of the Z-axis trajectory, the head rotation angle change data is extracted, and the angle difference of the facial region displacement trajectory of each frame is calculated. The posture transformation angle in each time series is paired with the corresponding feature point coordinate offset to obtain angle displacement paired data; S412: Based on the angle displacement paired data, detecting micro-motion speed information of each region, comparing the coordinate change amplitude and time interval of each region, and calculating the rate joint offset trend; S413: Based on the rate joint offset trend, perform data comparison on each group of micro-motion data and the spatial motion changes of the feature points, analyze the correlation between the structural similarity parameters and the motion linkage indicators in the combination, collect data on the combinations that meet the linkage conditions, and obtain motion linkage matching data.

7. The AI-based face recognition security method according to claim 1, characterized in that: The steps also include: S5: Based on the action linkage matching data, the correspondence between the numbered feature points and the behavior parameters is analyzed, the consistency of the data structure of each group of recognition channels is determined, the joint recognition integrated port discrimination process is optimized, the group with both structure and behavior meeting the standards is identified, the identification number is marked, and the face recognition mapping result is obtained; The face recognition mapping result includes an identity mapping identifier, an identification identifier, and a feature linkage mapping index.

8. The AI-based face recognition security method according to claim 7, characterized in that: The steps for obtaining the face recognition mapping result are specifically as follows: S511: Based on the action linkage matching data, analyzing the correspondence between the numbered feature points and the behavior parameters, determining the synchronization between the continuous distribution of each feature point in the three-dimensional space and the behavior change, screening corresponding groups with highly consistent structural features and behavioral features, and obtaining corresponding distribution data; S512: Based on the corresponding distribution data, compare the spatial structure relationship and behavior change pattern of the difference data group, optimize the correspondence rules between the arrangement order of the feature point combination and the behavior parameters, select the combination with consistent structure and behavior, and obtain structure mapping combination data; S513: Based on the structure mapping combination data, determine the linkage distribution between the feature point group and the behavior parameter group, calculate the participation ratio of each group of feature points and the linkage characteristics of the behavior parameters, mark the synchronization relationship and mapping index, and obtain the face recognition mapping result.

Images