Suspicious person monitoring method and system for bank security

Abstract

The application discloses a suspicious person monitoring method and system for bank security, and relates to the field of image processing. The method comprises the following steps: a multispectral imaging device collects personnel activity image data flow to construct a three-dimensional behavior image data body; a virtual behavior simulation environment is constructed, a multi-path standard behavior trajectory set is generated, and a behavior inertia benchmark atlas is constructed; cross-time domain optical flow consistency analysis and local visual structure stability calculation of the data body are performed to construct a behavior trajectory disturbance field distribution, a behavior inertia damage degree distribution matrix is generated based on a differential mapping relationship with the atlas; micro-scale visual enhancement processing is performed, a micro-motion tension tensor is constructed by using sub-pixel level jitter evolution detection and skin surface reflection dynamic fluctuation analysis; and the tension tensor and the damage degree distribution matrix are input into a visual risk time sequence evolution channel to generate a suspicious risk evolution curve. The application solves the technical problem of limited recognition accuracy in existing suspicious person monitoring, and achieves the technical effect of improving recognition accuracy.

Description

Technical Field

[0001] This application relates to the field of image processing, and in particular to methods and systems for detecting suspicious persons for bank security. Background Technology

[0002] Accurate monitoring of suspicious individuals in bank security is a core element in ensuring the operational safety of financial institutions and preventing malicious incidents, directly impacting the security of funds and the rights of personnel. Currently, the industry mainly relies on traditional video surveillance combined with manual patrols, or simple behavioral recognition algorithms based on single visible light images to screen suspicious individuals. These methods are prone to missed or false positives due to their over-reliance on subjective human judgment, and single visual data is insufficient to capture deep features such as micro-movements and behavioral inertia, resulting in low accuracy and delayed response in identifying concealed suspicious behavior.

[0003] Currently, the technology used for suspicious person monitoring in bank security suffers from limited accuracy. Summary of the Invention

[0004] This application provides a method and system for monitoring suspicious persons in bank security. It collects continuous time-series image data streams of personnel activities using multispectral imaging devices deployed in the bank's business area. This data is used to construct a three-dimensional behavioral image data volume containing temporal, spatial, and human skeleton posture dimensions. The system reads the bank's spatial structure, business equipment layout parameters, and historical normal business samples to build a virtual behavioral simulation environment. This generates a set of multi-path standard behavioral trajectories and constructs a behavioral inertial reference map. The system then performs cross-temporal optical flow consistency constraint analysis and local visual structure stability calculations on the three-dimensional behavioral image data volume to construct the behavioral trajectory perturbation of the target personnel. By combining the field distribution with the differential mapping relationship between it and the behavioral inertia baseline map to generate a behavioral inertia damage distribution matrix, the target personnel area is located and micro-scale visual enhancement is performed on the three-dimensional behavioral image data volume. Through sub-pixel level jitter evolution detection of key human skeleton points and dynamic fluctuation analysis of skin surface reflection, a micro-motion tension tensor is constructed. The micro-motion tension tensor and the behavioral inertia damage distribution matrix are input into the visual risk temporal evolution channel to generate a suspicious risk evolution curve. These technical means solve the technical problem of limited identification accuracy in existing suspicious personnel monitoring for bank security, and achieve the technical effect of improving the identification accuracy of security monitoring.

[0005] This application provides a method for monitoring suspicious persons in bank security, comprising: acquiring continuous temporal image data streams of personnel activities based on multispectral imaging devices deployed in the bank's business area, and constructing a three-dimensional behavioral image data volume including time dimension, spatial dimension, and human skeleton posture dimension; reading the bank's business space structure parameters, business equipment spatial layout parameters, and historical normal business samples to construct a virtual behavioral simulation environment to generate a set of multi-path standard behavioral trajectories and construct a behavioral inertial reference map; performing cross-temporal optical flow consistency constraint analysis and local visual structure stability calculation on the three-dimensional behavioral image data volume to construct the perturbation field distribution of the target person's behavioral trajectory, and generating a behavioral inertial destructive degree distribution matrix based on the differential mapping relationship with the behavioral inertial reference map; performing micro-scale visual enhancement processing on the three-dimensional behavioral image data volume after performing regional localization of the target person, and constructing a micro-motion tension tensor using sub-pixel level jitter evolution detection of key human skeleton points and dynamic fluctuation analysis of skin surface reflection; inputting the micro-motion tension tensor and the behavioral inertial destructive degree distribution matrix into a visual risk temporal evolution channel to generate a suspicious risk evolution curve.

[0006] In a possible implementation, a micro-motion tension tensor is constructed, and the following processing is performed: After completing the regional localization of the target person, the temporal pixel grayscale fluctuation sequence of the target region is extracted from the three-dimensional behavioral image data volume to construct the spatiotemporal grayscale perturbation volume of the target region; sub-pixel-level phase difference reconstruction processing is performed on the spatiotemporal grayscale perturbation volume to generate a micro-displacement phase field of the key skeleton point region; frequency domain energy spectrum decomposition is performed on the target region based on the micro-displacement phase field to generate a micro-motion high-frequency response texture map, and the skeleton point jitter evolution structure map is constructed using the micro-motion high-frequency response texture map; multispectral reflectance component separation processing is performed on the skin surface reflectance features of the target region to construct a skin surface reflectance dynamic change image sequence; tensor fusion encoding is performed on the skeleton point jitter evolution structure map and the skin surface reflectance dynamic change image sequence to construct the micro-motion tension tensor.

[0007] In a possible implementation, tensor fusion encoding is performed on the skeleton point jitter evolution structure map and the skin surface reflection dynamic change image sequence to construct a micro-motion tension tensor. The following processing is then performed: the temporal pixel grayscale fluctuation sequence is grouped into phase consistency groups according to a preset sub-pixel spatial sampling rule to construct pixel phase clustering units; in the pixel phase clustering units, phase stable structure units of skeleton point regions are constructed based on the spatial projection relationship of key human skeleton points, and the phase stable structure units are superimposed on the micro-displacement phase field; based on the micro-displacement phase field of the superimposed phase stable structure units, frequency domain energy spatial interleaving encoding is performed to generate the texture response density structure of the micro-motion high-frequency response texture map; the reflection phase temporal consistency correction is performed on the multispectral reflection component separation processing result to generate the reflection consistency structure of the skin surface reflection dynamic change image sequence; based on the tensor alignment rule between the texture response density structure and the reflection consistency structure, tensor fusion refinement encoding is performed on the micro-motion high-frequency response texture map and the skin surface reflection dynamic change image sequence to construct a micro-motion tension tensor.

[0008] In a possible implementation, a behavioral inertial reference map is constructed, and the following processing is performed: Discrete-time structured encoding is performed on each standard behavioral trajectory in the multi-path standard behavioral trajectory set to form a trajectory temporal structure unit, which is then used as the temporal structure component of the behavioral inertial reference map; spatial path topology normalization is performed on the standard behavioral trajectory set to form a trajectory spatial topology structure unit, which is then used as the spatial structure component of the inertial reference map; directional stability joint modeling is performed on the trajectory temporal structure unit and the trajectory spatial topology structure unit to generate a directional stability constraint structure, which is then solidified as the directional stability constraint component of the behavioral inertial reference map; the temporal structure component, spatial structure component, and directional stability constraint component are then fused and encoded in three dimensions to construct a behavioral inertial reference map with fixed inter-layer mapping rules.

[0009] In a possible implementation, the behavioral trajectory perturbation field distribution of the target person is constructed. Based on the differential mapping relationship with the behavioral inertial reference map, a behavioral inertial destructiveness distribution matrix is ​​generated, and the following processing is performed: a cross-temporal optical flow consistency constraint kernel is introduced into the three-dimensional behavioral image data volume to construct an optical flow constraint consistency tensor field, and the optical flow constraint consistency tensor field is used as the basic structural layer of the behavioral trajectory perturbation field; the visual structural stability distribution features of continuous image frames are executed to construct a local structural stability weight field, and the local structural stability weight field is used as the stability modulation layer of the behavioral trajectory perturbation field; the optical flow constraint consistency tensor field and the local structural stability weight field are coupled and superimposed between layers to construct a behavioral trajectory perturbation field distribution with propagation rules; the behavioral trajectory perturbation field distribution and the behavioral inertial reference map are scaled differentially mapped and encoded to construct a behavioral inertial destructiveness distribution matrix containing spatial hierarchical diffusion characteristics.

[0010] In a possible implementation, the behavioral trajectory perturbation field distribution and the behavioral inertial reference map are scale-difference mapped and encoded to construct a behavioral inertial destructiveness distribution matrix containing spatial hierarchical diffusion characteristics. The following processing is then performed: Based on the inter-layer mapping rules of the behavioral inertial reference map, a multi-scale grid alignment structure is constructed, and the behavioral trajectory perturbation field distribution is mapped to the corresponding scale space of the multi-scale grid alignment structure; a propagation memory weight structure is introduced in each scale space to encode the historical perturbation residue superposition of the behavioral trajectory perturbation field distribution, generating a scale perturbation accumulation structure; based on the structural difference relationship between the scale perturbation accumulation structure and the behavioral inertial reference map, a scale difference coupling matrix is ​​constructed; and the scale difference coupling matrix is ​​spatially hierarchically diffused and encoded to construct the behavioral inertial destructiveness distribution matrix.

[0011] In a possible implementation, a suspicious risk evolution curve is generated, and the following processing is performed: anomaly warning trigger analysis is performed based on the suspicious risk evolution curve, and a warning response is constructed; warning dispatch management is performed based on the warning response, the warning dispatch management including sending warning information upwards and real-time warning.

[0012] In a possible implementation, the abnormal warning trigger analysis is performed based on the suspected risk evolution curve, and the following processing is also performed: if the warning response reporting threshold is not reached, a monitoring instruction is established based on the suspected risk evolution curve; and the multispectral imaging device is controlled to perform monitoring and management of the target personnel based on the monitoring instruction.

[0013] In a possible implementation, a suspicious risk evolution curve is generated, and the following processing is also performed: additional data collection of the target personnel is performed, including voice data and network identification data; joint authentication of the suspicious risk evolution curve is performed based on the additional data; and monitoring and management are carried out based on the joint authentication results.

[0014] This application also provides a suspicious person monitoring system for bank security, including: a personnel activity image data stream acquisition module, used to acquire continuous temporal personnel activity image data streams based on multispectral imaging devices deployed in the bank's business area, and construct a three-dimensional behavioral image data volume including time dimension, spatial dimension, and human skeleton posture dimension; a behavioral inertial reference map construction module, used to read the structural parameters of the bank's business space, the spatial layout parameters of business equipment, and historical normal business samples, construct a virtual behavioral simulation environment to generate a set of multi-path standard behavioral trajectories, and construct a behavioral inertial reference map; and a differential mapping module, used to perform cross-temporal optical flow consistency of the three-dimensional behavioral image data volume. The system employs a constraint analysis and local visual structure stability calculation to construct the perturbation field distribution of the target person's behavioral trajectory. Based on the differential mapping relationship with the behavioral inertia baseline map, it generates a behavioral inertia destruction degree distribution matrix. A micro-motion tension tensor construction module is used to perform micro-scale visual enhancement processing on the three-dimensional behavioral image data volume after performing regional localization of the target person. It utilizes sub-pixel-level jitter evolution detection of key human skeletal points and dynamic fluctuation analysis of skin surface reflection to construct a micro-motion tension tensor. A visual risk temporal evolution module is used to input the micro-motion tension tensor and the behavioral inertia destruction degree distribution matrix into the visual risk temporal evolution channel to generate a suspected risk evolution curve.

[0015] The proposed method and system for monitoring suspicious persons in bank security involves the following steps: First, a continuous time-series image data stream of personnel activities is collected using a multispectral imaging device deployed in the bank's business area. This data stream is used to construct a three-dimensional behavioral image data volume, including temporal, spatial, and human skeleton posture dimensions. Then, the structural parameters of the bank's business space, the spatial layout parameters of business equipment, and historical normal business samples are read to construct a virtual behavioral simulation environment. This environment generates a set of multi-path standard behavioral trajectories and constructs a behavioral inertial reference map. Next, cross-temporal optical flow consistency constraint analysis and local visual structure stability calculation are performed on the three-dimensional behavioral image data volume to construct the perturbation field distribution of the target person's behavioral trajectory. Based on the differential mapping relationship with the behavioral inertial reference map, a behavioral inertial destructiveness distribution matrix is ​​generated. Then, after performing regional localization of the target person on the three-dimensional behavioral image data volume, micro-scale visual enhancement processing is applied. Subpixel-level jitter evolution detection of key human skeleton points and dynamic fluctuation analysis of skin surface reflection are used to construct a micro-motion tension tensor. Finally, the micro-motion tension tensor and the behavioral inertial destructiveness distribution matrix are input into a visual risk temporal evolution channel to generate a suspicious risk evolution curve. Through the above process, the method and system proposed in this application achieve the technical effect of improving the accuracy of security monitoring and identification. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a flowchart illustrating a method for monitoring suspicious persons for bank security, provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of a suspicious person monitoring system for bank security provided in an embodiment of this application.

[0019] Figure labeling: Personnel activity image data stream acquisition module 10, behavioral inertia baseline map construction module 20, differential mapping module 30, micro-motion tension tensor construction module 40, visual risk temporal evolution module 50. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] This application provides a method for monitoring suspicious persons in bank security, such as... Figure 1 As shown, the method includes: Step S100: Based on the multispectral imaging device deployed in the bank's business area, a continuous time-series image data stream of human activity is collected, and a three-dimensional behavioral image data volume including the time dimension, spatial dimension, and human skeleton posture dimension is constructed.

[0022] Specifically, multispectral imaging devices, including visible light cameras, infrared thermal imaging cameras, and near-infrared cameras, are deployed at key locations such as the bank lobby entrance, self-service equipment area, counters, and cash handling area. The frame rate of these imaging devices is set to 15 to 30 frames per second to ensure continuous movement changes of people. Through synchronous triggering of the multispectral imaging devices, images are acquired at the same timestamp from different locations and with different spectral types, forming a continuous time-series data stream of human activity images. For each acquired frame, a human skeleton detection algorithm is used to extract the two-dimensional coordinates of key skeletal points such as the head, neck, shoulders, elbows, wrists, hips, knees, and ankles. The spatial coordinate information of each frame is used as spatial dimension data, the timestamp information of consecutive frames is used as temporal dimension data, and the coordinates of the human skeleton points in each frame and the connections between these points are used as human skeleton pose dimension data. Based on the three-dimensional structure of time, space, and human skeleton pose, the image data stream is stored in a structured manner to construct a three-dimensional behavioral image data volume. The length of the time dimension is the number of continuously acquired image frames, the length of the space dimension is the pixel resolution of the imaging device, and the length of the human skeleton pose dimension is the number of extracted human key skeleton points.

[0023] Step S200: Read the structural parameters of the bank's business space, the spatial layout parameters of the business equipment, and historical normal business samples to construct a virtual behavior simulation environment, generate a set of multi-path standard behavior trajectories, and construct a behavior inertia benchmark map.

[0024] Specifically, the system uses Building Information Modeling (BIM) to retrieve the length, width, and height dimensions of the bank's business space, as well as the location parameters of walls and columns, and other structural parameters. It also retrieves spatial layout parameters, such as spatial coordinates and dimensions, of service equipment like teller counters, ATMs, queuing systems, and waiting seats, from an equipment ledger. Video data of personnel activity during bank business hours over the past year is collected, and videos of personnel engaging in normal transactions such as deposits, withdrawals, transfers, and inquiries are selected as historical normal business samples. Behavioral trajectory data of the personnel in these samples is extracted. Using a 3D simulation engine, the retrieved structural parameters of the business space and the spatial layout parameters of the service equipment are input to construct a virtual simulation environment consistent with the actual bank business area. The behavioral trajectory data from the historical normal business samples is then input into the virtual simulation environment. A trajectory generation algorithm simulates the behavioral trajectories of different business types and different personnel movement paths, generating a set of multi-path standard behavioral trajectories. Each standard behavioral trajectory includes a complete spatial coordinate sequence and a time sequence from the moment a person enters the business hall to the moment they leave. Based on the generated set of multi-path standard behavioral trajectories, the temporal, spatial, and directional characteristics of the trajectories are extracted to construct a behavioral inertia baseline map.

[0025] In one possible implementation, a behavioral inertia baseline map is constructed. Step S200 further includes step S210, which involves performing discrete-time structured encoding on each standard behavioral trajectory in the multi-path standard behavioral trajectory set to form trajectory temporal structure units. These trajectory temporal structure units are then used as the temporal structure components of the behavioral inertia baseline map. Specifically, a discrete time interval is set, for example, every 0.5 seconds is a discrete time point. The time series of each standard behavioral trajectory is discretized to obtain the spatial coordinates of the person corresponding to each discrete time point. The discretized time points are then serialized and encoded, and a unique temporal encoding value is assigned to each discrete time point according to the chronological order. The number of discrete time points, the temporal encoding values, and the spatial coordinates of the corresponding time points for each standard behavioral trajectory are structurally integrated to form trajectory temporal structure units. Each trajectory temporal structure unit is a two-dimensional array, where the number of rows is the number of discrete time points, and the number of columns is the sum of the dimensions of the temporal encoding values ​​and the spatial coordinates. The trajectory temporal structure units corresponding to all standard behavioral trajectories are summarized as the temporal structure components of the behavioral inertia baseline map.

[0026] For example, by setting the discrete time interval to 0.5 seconds, a standard behavioral trajectory with a duration of 20 seconds is discretized to obtain 40 discrete time points. Each time point is assigned a time sequence code value from 1 to 40, and the spatial coordinates of the person on the x-axis, y-axis, and z-axis corresponding to each time point are recorded to form a 40-row, 4-column two-dimensional array as the trajectory temporal structure unit of the trajectory. The trajectory temporal structure units of 2000 standard behavioral trajectories are summarized to form the temporal structure component of the behavioral inertia reference map.

[0027] Step S220 involves performing spatial path topology normalization on the standard behavioral trajectory set to form trajectory spatial topology units, which are then used as spatial structural components of the inertial reference map. Specifically, a unified three-dimensional rectangular coordinate system is established with the geometric center point of the bank's business area as the origin, and the spatial coordinates of all trajectories in the standard behavioral trajectory set are transformed to this unified coordinate system. For each transformed standard behavioral trajectory, key path nodes such as the starting point, ending point, and inflection points are extracted, and the spatial distance and relative orientation between key path nodes are calculated. A topology normalization algorithm is used to normalize the number of key path nodes, the proportion of spatial distances between nodes, and the relative orientation relationships of each trajectory, eliminating the influence of differences in walking speed and path length between different trajectories. The normalized key path node information and the topological relationships between nodes are then structurally integrated to form trajectory spatial topology units, where each trajectory spatial topology unit is a structure containing a set of nodes and a set of topological relationships. The trajectory spatial topology units corresponding to all standard behavioral trajectories are summarized as spatial structural components of the behavioral inertial reference map.

[0028] For example, a three-dimensional coordinate system is established with the geometric center point of a bank branch as the origin. The spatial coordinates of 2,000 standard behavioral trajectories are transformed into this coordinate system. Key nodes such as the entrance start point, ATM inflection point, and exit end point of each trajectory are extracted. The spatial distance ratio between nodes is calculated, and nodes with a distance ratio greater than 0.1 are retained. The retained nodes and the orientation relationships between nodes are normalized to form trajectory spatial topology units. All units are summarized to form the spatial structure components of the behavioral inertial reference map.

[0029] Step S230: Perform joint modeling of directional stability on the trajectory temporal structure unit and the trajectory spatial topology structure unit to generate a directional stability constraint structure, and solidify the directional stability constraint structure into directional stability constraint components of the behavioral inertia reference map. Specifically, for each discrete time point in the trajectory temporal structure unit, calculate the personnel movement direction vector corresponding to that time point. The direction vector is calculated as the spatial coordinate difference between the current time point and the next time point. For each critical path node in the trajectory spatial topology structure unit, calculate the personnel movement direction vector at that node. The direction vector is calculated as the spatial coordinate difference between the current node and the next node. Statistically analyze the distribution of direction vectors of all standard behavioral trajectories at the same discrete time point and the same critical path node, and calculate the mean and variance of the direction vectors. The mean represents the standard movement direction at that position, and the variance represents the stability of the direction. Based on the mean and variance of the direction vectors, construct a directional stability constraint model, and set a directional stability threshold. When the variance of the direction vector at a certain position is less than the threshold, the movement direction at that position is determined to be stable; otherwise, it is determined to be unstable. The parameters such as the directional mean, variance, and stability threshold in the directional stability constraint model are structurally integrated to generate a directional stability constraint structure. This directional stability constraint structure is then solidified into the directional stability constraint components of the behavioral inertial reference map.

[0030] For example, calculate the direction vectors of 2000 standard behavioral trajectories at discrete time points in front of an ATM. The mean of the direction vectors is along the positive x-axis, and the variance is 0.05. Set the direction stability threshold to 0.1. Since the variance is less than the threshold, the direction at this position is determined to be stable. Integrate the direction mean, variance, threshold, and other parameters at this position into a direction stability constraint structure. Summarize the constraint structures of all positions to form the direction stability constraint components of the behavioral inertia reference map.

[0031] Step S240 involves performing three-dimensional structural fusion encoding on the temporal structure component, spatial structure component, and directional stability constraint component to construct a behavioral inertia baseline map with fixed inter-layer mapping rules. Specifically, inter-layer mapping rules are set for the temporal structure component, spatial structure component, and directional stability constraint component. The discrete time points of the temporal structure component correspond one-to-one with the critical path nodes of the spatial structure component, and the temporal sequence encoding value of the temporal structure component corresponds one-to-one with the direction vector timestamp of the directional stability constraint component. A three-dimensional fusion encoding algorithm is used to fuse the data of the three components according to the set inter-layer mapping rules. During fusion, different weights are assigned to each component; for example, the weight of the temporal structure component is 0.4, the weight of the spatial structure component is 0.4, and the weight of the directional stability constraint component is 0.2. The fused data is stored in the form of a three-dimensional matrix, where the three dimensions of the matrix correspond to the time dimension, spatial dimension, and directional dimension, respectively. Each element in the matrix is ​​the fusion value of the three components in the corresponding dimension. The fused 3D matrix is ​​structured to form a behavioral inertia reference map with fixed inter-layer mapping rules. The inter-layer mapping rules are stored in the header information of the map in the form of a mapping table.

[0032] Step S300: Perform cross-temporal optical flow consistency constraint analysis and local visual structure stability calculation of the three-dimensional behavioral image data volume, construct the behavioral trajectory disturbance field distribution of the target person, and generate the behavioral inertia destruction degree distribution matrix based on the differential mapping relationship with the behavioral inertia reference map.

[0033] Specifically, for consecutive temporal image frames in a 3D behavioral image data volume, an optical flow calculation algorithm is used to calculate the optical flow field between adjacent frames, obtaining the movement velocity and direction of each pixel. Cross-temporal optical flow consistency constraints are set; for example, the change in the optical flow field between five adjacent frames should be less than 10%. Consistency analysis is performed on the calculated optical flow field to filter out regions that meet the constraints. For local visual structures in the 3D behavioral image data volume, such as pixel regions around human skeleton points, feature values ​​such as gray-level variance and edge density are calculated. The stability of the local visual structure is calculated based on the change in these feature values; the smaller the change in feature values, the higher the stability. The cross-temporal optical flow consistency analysis results and the local visual structure stability calculation results are fused to construct the perturbation field distribution of the target person's behavioral trajectory. The intensity of the perturbation field is positively correlated with the change in the optical flow field and negatively correlated with the stability of the local visual structure. The perturbation field distribution of the behavior trajectory is calculated by difference with the behavior inertia reference map. The difference between the perturbation field distribution and the corresponding position in the reference map is calculated. Based on the distribution of the difference, a behavior inertia damage distribution matrix is ​​generated. Each element in the matrix represents the degree of behavior inertia damage at the corresponding position.

[0034] In one possible implementation, a perturbation field distribution of the target person's behavioral trajectory is constructed. Based on the differential mapping relationship with the behavioral inertia reference map, a behavioral inertia disruption degree distribution matrix is ​​generated. Step S300 further includes step S310, introducing a cross-temporal optical flow consistency constraint kernel into the three-dimensional behavioral image data volume, constructing an optical flow constraint consistency tensor field, and using the optical flow constraint consistency tensor field as the basic structural layer of the behavioral trajectory perturbation field. Specifically, a cross-temporal optical flow consistency constraint kernel is constructed. The parameters of the constraint kernel include the temporal window size and the optical flow consistency threshold. For example, the temporal window size is 5 frames, and the optical flow consistency threshold is 10%. The three-dimensional behavioral image data volume is divided into blocks according to the temporal window size to obtain multiple consecutive image frame blocks. For each image frame block, the optical flow field between two adjacent frames within the block is calculated, and the optical flow change amplitude of each pixel within the temporal window is statistically analyzed. Pixels with optical flow change amplitudes less than the optical flow consistency threshold are marked as consistent pixels, and those greater than the threshold are marked as inconsistent pixels. An optical flow-consistent consistency tensor field is constructed based on the labeling results. Each tensor component of the tensor field contains the coordinates of the pixel, the amplitude of optical flow change, and consistency labeling information. The constructed optical flow-consistent consistency tensor field is used as the basic structural layer of the behavior trajectory perturbation field. In this field, the intensity of the tensor component corresponding to a consistent pixel is zero, while the intensity of the tensor component corresponding to a non-consistent pixel is positively correlated with the amplitude of optical flow change.

[0035] Step S320: Execute the visual structural stability distribution features of consecutive image frames to construct a local structural stability weight field. Use this local structural stability weight field as the stability modulation layer of the behavior trajectory perturbation field. Specifically, define the analysis region of the local visual structure, for example, a 30×30 pixel rectangular region centered on each key skeletal point of the human body. For consecutive image frames in the 3D behavior image data volume, extract visual structural features such as the gray-level mean, gray-level variance, and edge gradient of each analysis region. Calculate the visual structural feature difference of the same analysis region in consecutive image frames. Calculate the visual structural stability of the region based on the difference; the smaller the feature difference, the higher the stability. Based on the calculation results of visual structural stability, assign corresponding weight values ​​to each analysis region; the higher the stability, the smaller the weight value, and vice versa. Arrange the weight values ​​of all analysis regions in a structured manner according to their spatial coordinates to construct the local structural stability weight field. Use this local structural stability weight field as the stability modulation layer of the behavior trajectory perturbation field to modulate the tensor component intensity of the basic structural layer. The modulation method is to multiply the tensor component intensity of the basic structural layer by the weight value at the corresponding position.

[0036] For example, a 30×30 pixel analysis region centered on 17 key skeletal points of the human body is set, the gray-level variance of each region is extracted from 10 consecutive frames of images, the gray-level variance difference between adjacent frames is calculated, and stability is calculated based on the difference. Regions with high stability are assigned a weight value of 0.1, and regions with low stability are assigned a weight value of 0.9. The weight values ​​of all regions are arranged according to spatial coordinates to construct a local structural stability weight field, which serves as a stability modulation layer for the behavioral trajectory perturbation field.

[0037] Step S330: The optical flow constraint consistency tensor field and the local structural stability weight field are coupled and superimposed interlayer to construct a behavior trajectory perturbation field distribution with propagation rules. Specifically, an interlayer coupling and superposition rule is set, which involves multiplying the intensity of each tensor component of the optical flow constraint consistency tensor field by the weight value at the corresponding position in the local structural stability weight field to obtain the coupled tensor component intensity. According to the set coupling and superposition rule, the optical flow constraint consistency tensor field and the local structural stability weight field are coupled pixel-by-pixel to obtain the coupled tensor field. A propagation rule for the behavior trajectory perturbation is set, which includes a propagation direction from non-consistent pixels to surrounding consistent pixels, and a propagation intensity that decreases with increasing propagation distance. Based on the propagation rule, the coupled tensor field is processed to calculate the degree of influence of the perturbation of surrounding pixels on each pixel, and this degree of influence is superimposed on the coupled tensor component intensity. The propagated tensor field is then stored in a structured manner to construct a behavior trajectory perturbation field distribution with propagation rules.

[0038] Step S340 involves performing scale-difference mapping encoding on the behavioral trajectory perturbation field distribution and the behavioral inertial reference map to construct a behavioral inertial destructiveness distribution matrix containing spatial hierarchical diffusion characteristics. Specifically, a multi-scale grid alignment system is built based on the inter-layer mapping rules of the behavioral inertial reference map to match the behavioral trajectory perturbation field distribution to the corresponding scale level. Simultaneously, historical perturbation residual information is superimposed using a propagation memory weighting mechanism to form the perturbation accumulation result at each scale. This perturbation accumulation result is then differentially processed with the corresponding scale data of the behavioral inertial reference map to obtain a scale-difference coupling matrix. A spatial hierarchical diffusion rule is set, specifying the starting scale, step size, and attenuation coefficient of the diffusion. Based on this rule, the differential information is progressively transmitted between different scales, with the information intensity gradually attenuating as the scale level decreases. The differential data at each scale after hierarchical diffusion are normalized, mapping the values ​​to the interval between 0 and 1. Data from all scales are then integrated according to spatial location to finally construct a behavioral inertial destructiveness distribution matrix with spatial hierarchical diffusion characteristics.

[0039] In one possible implementation, the behavioral trajectory perturbation field distribution and the behavioral inertial reference map are encoded using scale difference mapping to construct a behavioral inertial destructive degree distribution matrix containing spatial hierarchical diffusion characteristics. Step S340 further includes step S341, which constructs a multi-scale grid alignment structure based on the inter-layer mapping rules of the behavioral inertial reference map, mapping the behavioral trajectory perturbation field distribution to the corresponding scale space of the multi-scale grid alignment structure. Specifically, the inter-layer mapping rules stored in the header of the behavioral inertial reference map are read, and the scale parameters of the time dimension, spatial dimension, and orientation dimension in the rules are extracted. Based on the extracted scale parameters, a multi-scale grid alignment structure is constructed, and the grid structure includes three scales: large, medium, and small. The behavioral trajectory perturbation field distribution is subjected to multi-scale downsampling and upsampling processing to generate perturbation field distribution data with the same resolution as the large, medium, and small scale grids, respectively. The perturbation field distribution data at different scales are mapped to the corresponding scale spaces according to the positional correspondence of the multi-scale raster alignment structure. Specifically, large-scale perturbation field data are mapped to large-scale raster space, medium-scale data are mapped to medium-scale space, and small-scale data are mapped to small-scale space, ensuring that the mapped data are spatially aligned with the corresponding scale data of the behavioral inertial reference map.

[0040] Step S342: Introduce a propagation memory weight structure in each scale space to perform historical perturbation residual superposition encoding on the behavior trajectory perturbation field distribution, generating a scale perturbation accumulation structure. Specifically, in each scale space of the multi-scale raster alignment structure, a propagation memory weight structure is constructed. The parameters of the weight structure include memory duration and memory decay coefficient, for example, the memory duration is 10 frames of images, and the memory decay coefficient is 0.9. For the behavior trajectory perturbation field distribution data in each scale space, the perturbation field data of the historical 10 frames of images are extracted in chronological order. Based on the parameters of the propagation memory weight structure, the historical perturbation field data is weighted, with the weight value being larger for historical frames closer to the current frame. For example, the weight value of the previous frame is 0.9, the weight value of the two previous frames is 0.81, and so on. The weighted historical perturbation field data is superimposed and encoded with the perturbation field data of the current frame using a weighted summation method to obtain the perturbation accumulation data in each scale space. The perturbation accumulation data in each scale space are structured and integrated to generate a scale perturbation accumulation structure, where the perturbation accumulation structure at each scale is a two-dimensional array containing time series and perturbation accumulation values.

[0041] Step S343: Based on the structural difference relationship between the scale perturbation accumulation structure and the behavioral inertial reference map, a scale difference coupling matrix is ​​constructed. Specifically, for each scale perturbation accumulation data in the scale perturbation accumulation structure, structural features of its time dimension, spatial dimension, and perturbation accumulation value are extracted. For the corresponding scale data in the behavioral inertial reference map, structural features of the same dimension are extracted. The structural difference between the scale perturbation accumulation data and the corresponding scale data in the behavioral inertial reference map is calculated. The difference calculation method includes the temporal difference in the time dimension, the coordinate difference in the spatial dimension, and the difference between the perturbation accumulation value and the reference value. Coupling rules are set, and the time dimension difference, spatial dimension difference, and perturbation accumulation value difference are coupled and calculated according to a preset weight ratio. The results after coupling calculation are arranged in scale order to construct a scale difference coupling matrix, where the number of rows in the matrix is ​​the number of scales, and the number of columns is the number of feature dimensions for each scale.

[0042] Step S344: Perform spatial hierarchical diffusion encoding on the scale difference coupling matrix to construct a behavioral inertial destructiveness distribution matrix. Specifically, set the parameters of the spatial hierarchical diffusion encoding, including the diffusion starting level, diffusion step size, and diffusion attenuation coefficient. For example, the diffusion starting level is large-scale space, the diffusion step size is 5 pixels, and the diffusion attenuation coefficient is 0.8. Using the large-scale difference data in the scale difference coupling matrix as the diffusion starting data, diffusion is performed hierarchically to the mesoscale and small-scale spaces according to the set diffusion step size and attenuation coefficient. During the diffusion process, the difference information of the large-scale data is transmitted to the mesoscale space, and the mesoscale data is then transmitted to the small-scale space. Each time it is transmitted, the intensity of the difference information is attenuated according to the attenuation coefficient. Normalize the diffusion of the difference data at each scale, mapping the data values ​​to the interval between 0 and 1. Integrate the normalized difference data at each scale according to spatial location to construct a behavioral inertial destructiveness distribution matrix. The resolution of the matrix is ​​consistent with the resolution of the largest scale space, and each element in the matrix represents the behavioral inertial destructiveness at the corresponding location.

[0043] Step S400: After performing regional localization of the target person on the three-dimensional behavioral image data volume, micro-scale visual enhancement processing is performed. Subpixel-level jitter evolution detection of key human skeletal points and dynamic fluctuation analysis of skin surface reflection are used to construct a micro-motion tension tensor.

[0044] Specifically, target detection algorithms, such as deep learning-based target detection models, are used to locate the target person region in each frame of the 3D behavioral image data volume, obtaining the bounding box coordinates of the target person. For the located target person region, micro-scale visual enhancement algorithms, such as histogram equalization and Laplacian enhancement algorithms, are used to enhance image details within the region and improve the recognizability of micro-scale features. For the enhanced image region, the coordinates of key human skeleton points are extracted, and sub-pixel-level detection algorithms, such as interpolation algorithms, are used to improve the coordinate accuracy of the skeleton points to the sub-pixel level. The sub-pixel-level skeleton point coordinates in consecutive image frames are tracked, and the jitter amplitude and frequency of the skeleton points between adjacent frames are calculated, performing sub-pixel-level jitter evolution detection. For the skin region of the target person, such as the face and hands, skin surface reflectance data collected by a multispectral imaging device is extracted, and the dynamic fluctuation amplitude and frequency of the reflectance data in consecutive image frames are calculated, performing dynamic fluctuation analysis of skin surface reflectance. The subpixel-level jitter evolution detection results and the skin surface reflection dynamic fluctuation analysis results are fused by tensor quantization to construct a micro-motion tension tensor. The tensor has dimensions including time dimension, skeleton point dimension, and reflection feature dimension.

[0045] In one possible implementation, a micro-motion tension tensor is constructed. Step S400 further includes step S410, whereby, after completing the regional localization of the target person, a temporal pixel grayscale fluctuation sequence of the target region is extracted from the three-dimensional behavioral image data volume to construct a spatiotemporal grayscale perturbation body of the target region. Specifically, based on the bounding box coordinates obtained from the regional localization of the target person, image data of the target region is extracted from each frame of the three-dimensional behavioral image data volume. For the extracted target region image data, the grayscale value of each pixel is extracted in chronological order to obtain a temporal grayscale value sequence of each pixel. The fluctuation amplitude of the temporal grayscale value sequence of each pixel is calculated, and the fluctuation amplitude is calculated as the difference between the maximum and minimum grayscale values ​​in the sequence. The spatial coordinates, temporal grayscale value sequence, and grayscale fluctuation amplitude of each pixel are structurally integrated to construct a spatiotemporal grayscale perturbation body of the target region according to the time dimension, spatial dimension, and grayscale fluctuation dimension. The length of the time dimension is the number of consecutively acquired image frames, the length of the spatial dimension is the number of pixels in the target region, and the length of the grayscale fluctuation dimension is the quantization level of the grayscale fluctuation amplitude.

[0046] Step S420: Perform sub-pixel-level phase difference reconstruction processing on the spatiotemporal grayscale perturbation to generate the micro-displacement phase field of the key skeleton point region. Specifically, the temporal grayscale value sequence in the spatiotemporal grayscale perturbation is transformed to the frequency domain using a Fourier transform algorithm to obtain the phase information and amplitude information of each pixel. A sub-pixel-level phase difference calculation rule is set, which calculates the phase difference value of corresponding pixels in two adjacent frames, with the calculation precision at the sub-pixel level. According to the set rule, phase difference calculation is performed on all pixels in the spatiotemporal grayscale perturbation to obtain the sub-pixel-level phase difference value sequence for each pixel. The pixel region surrounding the key skeleton points of the human body is extracted, for example, a 10×10 pixel region centered on each skeleton point. The sub-pixel-level phase difference value sequences of all pixels within this region are averaged to obtain the average phase difference value of the key skeleton point region. The spatial coordinates of key skeleton points and the corresponding average phase difference values ​​are structurally integrated to generate a micro-displacement phase field for the key skeleton point region. Each element of the micro-displacement phase field represents the sub-pixel level phase difference value of the corresponding skeleton point region.

[0047] Step S430: Based on the micro-displacement phase field, perform frequency domain energy spectrum decomposition on the target region to generate a micro-motion high-frequency response texture map. Use this texture map to construct a jitter evolution structure map of the skeleton points. Specifically, for the phase difference data in the micro-displacement phase field, perform frequency domain energy spectrum decomposition using a fast Fourier transform algorithm to obtain the energy distribution of different frequency bands. Set a high-frequency threshold, for example, frequencies greater than 10Hz are considered high-frequency bands, and extract high-frequency energy data from the frequency domain energy spectrum. Convert the high-frequency energy data back to the spatial domain to generate a micro-motion high-frequency response texture map. The pixel values ​​of the texture map are positively correlated with the high-frequency energy values; the higher the energy value, the larger the pixel value. Perform threshold segmentation on the micro-motion high-frequency response texture map to extract regions with energy values ​​greater than the set threshold. These regions correspond to the jitter regions of key human skeleton points. Based on the threshold segmentation results, the jitter position and jitter range of each key skeleton point are marked. The jitter positions and ranges in consecutive image frames are connected in chronological order to construct a jitter evolution structure diagram of the skeleton points. The structure diagram contains the jitter trajectory and jitter intensity information of each skeleton point at different time points.

[0048] Step S440 involves performing multispectral reflectance component separation processing on the skin surface reflectance features of the target area to construct a dynamic change image sequence of skin surface reflectance. Specifically, image data from different spectral channels acquired by a multispectral imaging device, including visible light channels, infrared channels, and near-infrared channels, are extracted from the target area of ​​the three-dimensional behavioral image data volume. For the image data of different spectral channels, a multispectral reflectance component separation algorithm, such as independent component analysis, is used to separate different reflectance components of skin surface reflectance, such as diffuse reflectance and specular reflectance. For each separated reflectance component, its reflectance intensity value in consecutive image frames is calculated to obtain a temporal reflectance intensity sequence for each reflectance component. Based on the temporal reflectance intensity sequence, the reflectance intensity difference between two adjacent image frames is calculated. According to the magnitude and sign of the difference, a dynamic change image of skin surface reflectance is generated, where the pixel value in the image represents the magnitude of the reflectance intensity change at the corresponding location. The dynamic change images of skin surface reflectance corresponding to consecutive image frames are arranged in chronological order to construct a dynamic change image sequence of skin surface reflectance.

[0049] Step S450: Perform tensor fusion encoding on the skeleton point jitter evolution structure diagram and the skin surface reflection dynamic change image sequence to construct a micro-motion tension tensor. Specifically, for the skeleton point jitter evolution structure diagram, extract features such as jitter amplitude, jitter frequency, and jitter trajectory length for each key skeleton point, and quantize and encode these features to obtain a skeleton point jitter feature vector. For the skin surface reflection dynamic change image sequence, extract features such as reflection intensity change amplitude, change frequency, and change area area for each image frame, and quantize and encode these features to obtain a skin surface reflection feature vector. Set tensor fusion encoding rules, which align the skeleton point jitter feature vector and the skin surface reflection feature vector according to the time dimension, and the alignment method is to fuse feature vectors under the same timestamp. According to the set rules, the two feature vectors are weighted and fused. The fusion weight is set according to the importance of the features. For example, the weight of the skeleton point jitter feature is 0.6, and the weight of the skin surface reflection feature is 0.4. Specifically, the weighted fusion calculation can be completed by the dot product of the feature vectors. The fused feature vectors are used to construct a three-dimensional tensor, namely the micro-motion tension tensor, according to the time dimension, skeleton point dimension, and reflection feature dimension. Each element in the tensor represents the micro-motion tension value in the corresponding dimension.

[0050] In one possible implementation, the skeleton point jitter evolution structure map and the skin surface reflection dynamic change image sequence are subjected to tensor fusion encoding to construct a micro-motion tension tensor. Step S450 further includes step S451, which involves performing phase consistency grouping on the temporal pixel grayscale fluctuation sequence according to a preset sub-pixel spatial sampling rule to construct a pixel phase clustering unit. Specifically, a preset sub-pixel spatial sampling rule is set, including a sampling interval, for example, a sampling interval of 0.5 pixels, and a sampling area of ​​all pixels within the target area. For each pixel in the temporal pixel grayscale fluctuation sequence, sub-pixel level sampling is performed according to the set sampling rule to obtain the phase information of each sampling point. The difference between the phase value of each sampling point and a preset reference phase value is calculated, and sampling points with a phase difference less than a preset threshold are grouped together, performing phase consistency grouping. The spatial coordinates, phase values, and phase differences of each group of sampling points are structurally integrated to construct a pixel phase clustering unit, wherein each pixel phase clustering unit is a structure containing a set of sampling points and phase features.

[0051] Step S452: In the pixel phase clustering unit, based on the spatial projection relationship of key human skeleton points, a phase-stabilized structural unit for the skeleton point region is constructed, and the phase-stabilized structural unit is superimposed on the micro-displacement phase field. Specifically, the spatial coordinates of the key human skeleton points are projected to obtain the projected coordinates of the skeleton points in the space where the pixel phase clustering unit is located. In the pixel phase clustering unit, a certain range of regions, such as a 5×5 pixel region, is defined as the skeleton point region, centered on the projected coordinates of the skeleton points. The mean and variance of the phase values ​​of all sampling points within the skeleton point region are calculated. The mean represents the standard phase value of the region, and the variance represents the stability of the phase. The spatial coordinates, mean phase, and variance of the skeleton point region are structurally integrated to construct a phase-stabilized structural unit for the skeleton point region. The mean phase and variance in the phase-stabilized structural unit are superimposed on the phase data of the corresponding skeleton point region in the micro-displacement phase field. The superposition method is to add the mean phase to the phase value of the micro-displacement phase field multiplied by a weighting coefficient.

[0052] Step S453: Based on the micro-displacement phase field of the superimposed phase-stabilized structural units, frequency domain energy spatial interleaving encoding is performed to generate the texture response density structure of the micro-motion high-frequency response texture map. Specifically, for the micro-displacement phase field data after superimposing phase-stabilized structural units, a two-dimensional Fourier transform algorithm is used to perform frequency domain energy decomposition to obtain the frequency domain energy distribution at different spatial locations. A high-frequency energy threshold is set, and high-frequency energy data greater than the threshold in the frequency domain energy distribution is extracted. Spatial interleaving encoding is performed on the high-frequency energy data, and the encoding method is to interleave the high-frequency energy data at different spatial locations in row and column order to form texture response density data. The texture response density data is structurally integrated according to spatial coordinates to generate the texture response density structure of the micro-motion high-frequency response texture map, wherein each element in the structure represents the texture response density value at the corresponding spatial location.

[0053] Step S454 involves performing temporal consistency correction on the multispectral reflectance component separation processing results to generate a reflectance consistency structure for the dynamic change image sequence of skin surface reflectance. Specifically, the temporal phase information of each component is extracted from the diffuse reflectance and specular reflectance component data obtained from the multispectral reflectance component separation processing. A temporal consistency correction rule for reflectance phase is set, which calculates the phase difference between consecutive image frames and corrects phase values ​​with differences greater than a preset threshold to the phase value of the previous frame. The temporal phase information of the diffuse reflectance and specular reflectance components is corrected according to the set correction rule to obtain a corrected phase sequence. Based on the corrected phase sequence, the dynamic change value of skin surface reflectance intensity is calculated to generate a corrected image sequence of dynamic change in skin surface reflectance. Reflectance consistency features of each image frame in the corrected image sequence are extracted, such as the variance of the reflectance phase and the standard deviation of the reflectance intensity. These features are then structurally integrated to generate a reflectance consistency structure for the dynamic change image sequence of skin surface reflectance.

[0054] Step S455: Based on the tensor alignment rules between the texture response density structure and the reflection consistency structure, tensor fusion and refinement encoding are performed on the micro-motion high-frequency response texture map and the skin surface reflection dynamic change image sequence to construct a micro-motion tension tensor. Specifically, a tensor alignment rule is set between the texture response density structure and the reflection consistency structure, which aligns the spatial dimension of the texture response density structure with the spatial dimension of the reflection consistency structure, and the temporal dimension with the temporal dimension. For the micro-motion high-frequency response texture map, texture response density feature vectors are extracted based on the texture response density structure; for the skin surface reflection dynamic change image sequence, reflection consistency feature vectors are extracted based on the reflection consistency structure. According to the tensor alignment rules, the texture response density feature vectors and reflection consistency feature vectors are dimensionally matched by expanding the lower-dimensional feature vectors to match the dimensions of the higher-dimensional feature vectors. The matched feature vectors are then refined and encoded by calculating the dot product of the two feature vectors to obtain a fused feature value. The fused feature value is then used to construct a three-dimensional tensor, namely the micro-motion tension tensor, according to the temporal dimension, spatial dimension, and feature dimension, completing the tensor fusion and refinement encoding.

[0055] Step S500: Input the micro-motion tension tensor and the behavioral inertia destruction distribution matrix into the visual risk time-series evolution channel to generate a suspected risk evolution curve.

[0056] Specifically, a visual risk temporal evolution channel is constructed, comprising three parts: a feature fusion layer, a temporal analysis layer, and a curve generation layer. The micro-motion tension tensor and the behavioral inertia disruption distribution matrix are input into the feature fusion layer. A feature concatenation algorithm is used to concatenate the feature dimensions of the two datasets, resulting in a fused feature matrix. This fused feature matrix is ​​then input into the temporal analysis layer, where a temporal convolutional network algorithm is used to extract temporal features, yielding risk feature values ​​at different time points. Finally, the risk feature values ​​output from the temporal analysis layer are input into the curve generation layer. Using time as the horizontal axis and risk feature values ​​as the vertical axis, a curve fitting algorithm, such as least squares, is employed to fit the risk feature values, generating a suspicious risk evolution curve. Each point on the curve represents the degree of suspicious risk at the corresponding time point.

[0057] In one possible implementation, a suspicious risk evolution curve is generated, and the method further includes step S600, performing anomaly warning trigger analysis based on the suspicious risk evolution curve, and constructing a warning response.

[0058] Specifically, anomaly warning trigger thresholds are set, including Level 1, Level 2, and Level 3 warning thresholds. The risk value at each time point in the suspected risk evolution curve is compared with the warning trigger threshold to determine whether the risk value reaches or exceeds a certain level of threshold. Based on the comparison results, anomaly warning trigger analysis is performed. If the risk value reaches the Level 1 threshold, it is determined as a low-risk warning; if it reaches the Level 2 threshold, it is determined as a medium-risk warning; and if it reaches the Level 3 threshold, it is determined as a high-risk warning. Warning responses are constructed according to the warning levels. The warning response content includes the warning level, trigger time, target personnel location information, and risk characteristic data. These contents are structured and integrated to form a standardized warning response message.

[0059] Step S700: Based on the warning response, perform early warning dispatch management, which includes sending early warning information upwards and real-time early warning.

[0060] Specifically, an early warning and dispatch management system is constructed, comprising an information sending module and an early warning module. For upward early warning information transmission, the information sending module reads information such as the warning level, trigger time, and target personnel location from the early warning response message, and sends the early warning information to the bank's security management platform according to a preset communication protocol, such as a transmission control protocol. The transmitted information is in a standardized text format. For real-time early warnings, the early warning module reads the warning level from the early warning response message and triggers different warning methods based on different levels. For example, a low-risk warning triggers a desktop pop-up warning, a medium-risk warning triggers a desktop pop-up and an audio warning, and a high-risk warning triggers a desktop pop-up, an audio warning, and an on-site audible and visual alarm. Simultaneously, the warning message displays the target personnel's location and risk level.

[0061] In one possible implementation, the method further includes step S800, which involves triggering anomaly warnings based on the suspected risk evolution curve. If the warning response reporting threshold is not reached, a monitoring instruction is established based on the suspected risk evolution curve.

[0062] Specifically, a lower limit is set for the warning response reporting threshold. The maximum risk value in the suspected risk evolution curve is compared with this lower limit. If the maximum risk value is less than the lower limit, it is determined that the warning response reporting threshold has not been reached. For suspected risk evolution curves that have not reached the reporting threshold, the risk value fluctuation characteristics in the curve are extracted, such as fluctuation frequency and fluctuation amplitude, to determine the time interval and behavioral characteristics of the target personnel that need to be monitored. Based on the extracted risk fluctuation characteristics and behavioral characteristics of the target personnel, a monitoring instruction is constructed. The instruction includes the regional location coordinates of the target personnel, the duration of monitoring, and the behavioral characteristics that need to be focused on, such as the amplitude of skeletal point jitter and the amplitude of skin surface reflection fluctuation.

[0063] Step S900: Control the multispectral imaging device to perform target personnel monitoring and management according to the monitoring instruction.

[0064] Specifically, the system reads information from the monitoring command, including the target person's location coordinates, monitoring duration, and key behavioral characteristics. Based on the target person's location coordinates, the system controls the multispectral imaging device deployed in that area to adjust its shooting parameters, such as increasing the frame rate to 30 frames per second and improving the image resolution to 4000×2000 pixels, to ensure clear capture of the target person's micro-movements. The monitoring duration is set, for example, 10 minutes, and the multispectral imaging device continuously acquires image data of the target person within that time. Key behavioral characteristics are extracted in real-time from the acquired image data, such as the amplitude of skeletal point jitter and skin surface reflection fluctuations, and compared with preset normal characteristic ranges. If the characteristic values ​​exceed the normal range, the priority of the monitoring command is updated. After the monitoring period ends, a monitoring report is generated, including the monitoring time, changes in the target person's behavioral characteristics, and whether any abnormal characteristics were observed.

[0065] In one possible implementation, a suspicious risk evolution curve is generated, and the method further includes step S1000, performing additional data collection on the target personnel, the additional data including voice data and network identification data.

[0066] Specifically, sound pickup devices, such as microphone arrays, are deployed in the bank's business area. The acquisition parameters of these devices are set, and they are controlled to collect voice data from the area where the target person is located. The collected voice data includes information such as the target person's speech content, tone, and speech rate. The target person's networked identification data is retrieved through the bank's networked identification system, including identity verification data, transaction record data, and historical security record data. Identity verification data is extracted from image data using facial recognition algorithms, transaction record data is retrieved from the bank's business system, and historical security record data is retrieved from the security management platform. The collected voice data undergoes noise reduction processing, for example, using adaptive filtering algorithms to remove environmental noise and improve the clarity of the voice data. The retrieved networked identification data is then structured and stored according to data type, forming an additional data set for the target person.

[0067] Step S1100: Perform joint authentication of the suspicious risk evolution curve based on the additional data, and conduct monitoring and management based on the joint authentication results.

[0068] Specifically, a joint authentication model is constructed. The model's inputs are the risk characteristic values ​​of the suspicious risk evolution curve and the characteristic values ​​of supplementary data. The output is the risk level after joint authentication. The risk characteristic values ​​of the suspicious risk evolution curve, such as risk peak and fluctuation frequency, and the characteristic values ​​of supplementary data, such as the probability of tense tone in voice data and the number of abnormal transactions in network identification data, are input into the joint authentication model. A weighted fusion algorithm is used to fuse the input characteristic values. The fusion weights are set according to the importance of the features; for example, the risk peak has a weight of 0.4, the probability of tense tone has a weight of 0.3, and the number of abnormal transactions has a weight of 0.3. The risk level after joint authentication is determined based on the fusion calculation results, and the levels include low risk, medium risk, and high risk. Monitoring and management are implemented according to the risk level of joint authentication. Low-risk levels continue to be monitored, medium-risk levels trigger a level-two warning response, and high-risk levels trigger a level-three warning response.

[0069] This application embodiment uses a multispectral imaging device deployed in the bank's business area to collect continuous temporal image data streams of personnel activities. This data is used to construct a three-dimensional behavioral image data volume containing temporal, spatial, and human skeleton posture dimensions. The system reads the bank's business space structure, business equipment layout parameters, and historical normal business samples to build a virtual behavioral simulation environment. It generates a set of multi-path standard behavioral trajectories and constructs a behavioral inertial reference map. Cross-temporal optical flow consistency constraint analysis and local visual structure stability calculation are performed on the three-dimensional behavioral image data volume to construct the perturbation field distribution of the target personnel's behavioral trajectory. Combined with the differential mapping relationship between this and the behavioral inertial reference map, a behavioral inertial destructiveness distribution matrix is ​​generated. The system then performs target personnel area localization and micro-scale visual enhancement on the three-dimensional behavioral image data volume. Through sub-pixel-level jitter evolution detection of key human skeleton points and dynamic fluctuation analysis of skin surface reflection, a micro-motion tension tensor is constructed. The micro-motion tension tensor and the behavioral inertial destructiveness distribution matrix are input into the visual risk temporal evolution channel to generate a suspicious risk evolution curve. These techniques solve the technical problem of limited identification accuracy in existing suspicious personnel monitoring for bank security, achieving the technical effect of improving the identification accuracy of security monitoring.

[0070] In the above text, refer to Figure 1 A method for detecting suspicious persons for bank security according to an embodiment of the present invention is described in detail. Next, reference will be made to... Figure 2 A suspicious person monitoring system for bank security is described according to an embodiment of the present invention.

[0071] The suspicious person monitoring system for bank security according to embodiments of the present invention addresses the technical problem of limited identification accuracy in existing suspicious person monitoring systems for bank security, thereby improving the technical effect of security monitoring and identification accuracy. The suspicious person monitoring system for bank security includes: a personnel activity image data stream acquisition module 10, a behavioral inertial reference map construction module 20, a differential mapping module 30, a micro-motion tension tensor construction module 40, and a visual risk temporal evolution module 50.

[0072] The personnel activity image data stream acquisition module 10 is used to acquire continuous temporal personnel activity image data streams based on multispectral imaging devices deployed in the bank's business area, and construct a three-dimensional behavioral image data volume including time dimension, spatial dimension, and human skeleton posture dimension; the behavior inertial reference map construction module 20 is used to read the bank's business space structure parameters, business equipment spatial layout parameters, and historical normal business samples to construct a virtual behavior simulation environment to generate a set of multi-path standard behavior trajectories and construct a behavior inertial reference map; the differential mapping module 30 is used to perform cross-temporal optical flow consistency constraint analysis and local visual structure of the three-dimensional behavioral image data volume. Stability calculation involves constructing the perturbation field distribution of the target person's behavioral trajectory and generating a behavioral inertia disruption distribution matrix based on the differential mapping relationship with the behavioral inertia baseline map. A micro-motion tension tensor construction module 40 is used to perform micro-scale visual enhancement processing on the three-dimensional behavioral image data volume after performing regional localization of the target person. It utilizes sub-pixel-level jitter evolution detection of key human skeletal points and dynamic fluctuation analysis of skin surface reflection to construct a micro-motion tension tensor. A visual risk temporal evolution module 50 is used to input the micro-motion tension tensor and the behavioral inertia disruption distribution matrix into the visual risk temporal evolution channel to generate a suspected risk evolution curve.

[0073] The detailed configuration of the micro-motion tension tensor construction module 40 is explained below: As mentioned above, the micro-motion tension tensor construction module 40 may further include: a temporal pixel grayscale fluctuation sequence extraction unit, used to extract the temporal pixel grayscale fluctuation sequence of the target region from the three-dimensional behavioral image data volume after completing the regional positioning of the target person, and construct the spatiotemporal grayscale perturbation body of the target region; and a sub-pixel level phase difference reconstruction processing unit, used to perform sub-pixel level phase difference reconstruction processing on the spatiotemporal grayscale perturbation body to generate key skeleton point regions. The micro-displacement phase field; the frequency domain energy spectrum decomposition unit is used to perform frequency domain energy spectrum decomposition on the target region based on the micro-displacement phase field to generate a micro-motion high-frequency response texture map, and to construct a skeleton point jitter evolution structure map using the micro-motion high-frequency response texture map; the multispectral reflectance component separation processing unit is used to perform multispectral reflectance component separation processing on the skin surface reflectance features of the target region to construct a skin surface reflectance dynamic change image sequence; the tensor fusion encoding unit is used to perform tensor fusion encoding on the skeleton point jitter evolution structure map and the skin surface reflectance dynamic change image sequence to construct a micro-motion tension tensor.

[0074] Specifically, the tensor fusion encoding of the skeleton point jitter evolution structure diagram and the skin surface reflection dynamic change image sequence is performed to construct a micro-motion tension tensor. The tensor fusion encoding unit may further include: a phase consistency grouping subunit for performing phase consistency grouping on the temporal pixel grayscale fluctuation sequence according to a preset sub-pixel spatial sampling rule to construct a pixel phase clustering unit; a phase stabilization structure unit construction subunit for constructing a phase stabilization structure unit for the skeleton point region in the pixel phase clustering unit based on the spatial projection relationship of key human skeleton points, and superimposing the phase stabilization structure unit onto the micro-displacement phase field; and a frequency domain energy space interleaving encoding subunit. The first unit is used to perform frequency domain energy space interleaving encoding on the micro-displacement phase field based on the superimposed phase stable structure unit to generate the texture response density structure of the micro-motion high-frequency response texture map; the second unit is used to perform reflection phase temporal consistency correction on the multispectral reflection component separation processing result to generate the reflection consistency structure of the skin surface reflection dynamic change image sequence; the third unit is used to perform tensor fusion and refinement encoding on the micro-motion high-frequency response texture map and the skin surface reflection dynamic change image sequence based on the tensor alignment rules between the texture response density structure and the reflection consistency structure to construct the micro-motion tension tensor.

[0075] The specific configuration of the behavior inertial reference map construction module 20 is described in detail below: As mentioned above, the behavior inertial reference map construction module 20 may further include: a discrete-time structured encoding unit for performing discrete-time structured encoding on each standard behavior trajectory in the multi-path standard behavior trajectory set to form a trajectory temporal structure unit, which is used as the temporal structure component of the behavior inertial reference map; a spatial path topology normalization processing unit for performing spatial path topology normalization processing on the standard behavior trajectory set to form a trajectory spatial topology structure unit, which is used as the spatial structure component of the inertial reference map; a directional stability joint modeling unit for performing directional stability joint modeling on the trajectory temporal structure unit and the trajectory spatial topology structure unit to generate a directional stability constraint structure, which is solidified as the directional stability constraint component of the behavior inertial reference map; and a three-dimensional structure fusion encoding unit for performing three-dimensional structure fusion encoding on the temporal structure component, spatial structure component, and directional stability constraint component to construct a behavior inertial reference map with fixed inter-layer mapping rules.

[0076] The detailed description of the specific configuration of the differential mapping module 30 is explained as follows: As mentioned above, the distribution of the target person's behavioral trajectory perturbation field is constructed. Based on the differential mapping relationship with the behavioral inertia reference map, a behavioral inertia destruction degree distribution matrix is ​​generated. The differential mapping module 30 may further include: an optical flow constraint consistency tensor field construction unit for introducing a cross-temporal optical flow consistency constraint kernel into the three-dimensional behavioral image data volume, constructing an optical flow constraint consistency tensor field, and using the optical flow constraint consistency tensor field as the basic structural layer of the behavioral trajectory perturbation field; a local structural stability weight field construction unit for... Based on the visual structural stability distribution characteristics of consecutive image frames, a local structural stability weight field is constructed, which is used as the stability modulation layer of the behavioral trajectory perturbation field. The interlayer coupling superposition unit is used to perform interlayer coupling superposition of the optical flow constraint consistency tensor field and the local structural stability weight field to construct a behavioral trajectory perturbation field distribution with propagation rules. The scale difference mapping encoding unit is used to perform scale difference mapping encoding on the behavioral trajectory perturbation field distribution and the behavioral inertia reference map to construct a behavioral inertia destruction degree distribution matrix containing spatial hierarchical diffusion characteristics.

[0077] Specifically, the behavioral trajectory perturbation field distribution and the behavioral inertial reference map are scale-difference mapped and encoded to construct a behavioral inertial destructiveness distribution matrix containing spatial hierarchical diffusion characteristics. The scale-difference mapping and encoding unit may further include: a multi-scale grid alignment structure construction subunit for constructing a multi-scale grid alignment structure based on the inter-layer mapping rules of the behavioral inertial reference map, mapping the behavioral trajectory perturbation field distribution to the corresponding scale space of the multi-scale grid alignment structure; a historical perturbation residue superposition encoding subunit for introducing a propagation memory weight structure in each scale space, performing historical perturbation residue superposition encoding on the behavioral trajectory perturbation field distribution to generate a scale perturbation accumulation structure; a scale-difference coupling matrix construction subunit for constructing a scale-difference coupling matrix based on the structural difference relationship between the scale perturbation accumulation structure and the behavioral inertial reference map; and a spatial hierarchical diffusion encoding subunit for performing spatial hierarchical diffusion encoding on the scale-difference coupling matrix to construct a behavioral inertial destructiveness distribution matrix.

[0078] The system, which generates a suspicious risk evolution curve, may further include: an anomaly warning trigger analysis module for performing anomaly warning trigger analysis based on the suspicious risk evolution curve and constructing a warning response; and a warning dispatch management module for performing warning dispatch management based on the warning response, wherein the warning dispatch management includes sending warning information upwards and real-time warnings.

[0079] The system, which triggers anomaly warnings based on the suspected risk evolution curve, may further include: a monitoring instruction establishment module for establishing a monitoring instruction based on the suspected risk evolution curve if the warning response reporting threshold is not reached; and a monitoring management module for controlling a multispectral imaging device to perform monitoring management of target personnel based on the monitoring instruction.

[0080] The system, which generates a suspicious risk evolution curve, may further include: an additional data acquisition module for collecting additional data on the target personnel, the additional data including voice data and network identification data; and a joint authentication module for performing joint authentication of the suspicious risk evolution curve based on the additional data, and for monitoring and management based on the joint authentication results.

[0081] The suspicious person monitoring system for bank security provided in this embodiment of the invention can execute the suspicious person monitoring method for bank security provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0082] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for monitoring suspicious persons in bank security, characterized in that, The method includes: Based on the continuous time-series image data stream of human activities collected by multispectral imaging devices deployed in the bank's business area, a three-dimensional behavioral image data volume including time dimension, spatial dimension and human skeleton posture dimension is constructed. Read the structural parameters of the bank's business space, the spatial layout parameters of business equipment, and historical normal business samples to construct a virtual behavior simulation environment, generate a set of multi-path standard behavior trajectories, and construct a behavior inertia benchmark map; Perform cross-temporal optical flow consistency constraint analysis and local visual structure stability calculation on the three-dimensional behavioral image data volume, construct the perturbation field distribution of the target person's behavioral trajectory, and generate the behavioral inertial damage distribution matrix based on the differential mapping relationship with the behavioral inertial reference map; After performing regional localization of the target person on the three-dimensional behavioral image data volume, micro-scale visual enhancement processing is performed. Sub-pixel level jitter evolution detection of key human skeletal points and dynamic fluctuation analysis of skin surface reflection are used to construct a micro-motion tension tensor. The micro-motion tension tensor and the behavioral inertia disruption distribution matrix are input into the visual risk time-series evolution channel to generate a suspected risk evolution curve. Constructing the micro-motion tension tensor includes: After completing the regional positioning of the target personnel, the temporal pixel grayscale fluctuation sequence of the target region is extracted from the three-dimensional behavioral image data volume to construct the spatiotemporal grayscale perturbation volume of the target region. The spatiotemporal grayscale perturbation body is subjected to subpixel-level phase difference reconstruction processing to generate a micro-displacement phase field in the key skeleton point region; Based on the micro-displacement phase field, frequency domain energy spectrum decomposition is performed on the target region to generate a micro-motion high-frequency response texture map, and the micro-motion high-frequency response texture map is used to construct a skeleton point jitter evolution structure map. Multispectral reflectance component separation processing is performed on the skin surface reflectance features of the target area to construct a dynamic image sequence of skin surface reflectance changes. The skeleton point jitter evolution structure diagram and the image sequence of skin surface reflection dynamic changes are subjected to tensor fusion encoding to construct the micro-motion tension tensor; The skeleton point jitter evolution structure diagram and the image sequence of dynamic changes in skin surface reflection are subjected to tensor fusion encoding to construct a micro-motion tension tensor, including: The temporal pixel grayscale fluctuation sequence is grouped into phase consistency groups according to a preset sub-pixel spatial sampling rule to construct a pixel phase clustering unit; In the pixel phase clustering unit, a phase-stabilized structure unit is constructed in the skeleton point region based on the spatial projection relationship of key human skeleton points, and the phase-stabilized structure unit is superimposed on the micro-displacement phase field. Based on the micro-displacement phase field of the superimposed phase-stabilized structural unit, frequency domain energy space interleaving coding is performed to generate the texture response density structure of the micro-motion high-frequency response texture map; The reflection phase temporal consistency is corrected for the multispectral reflectance component separation processing results to generate a reflectance consistency structure for the dynamic change image sequence of skin surface reflectance; Based on the tensor alignment rules between the texture response density structure and the reflection consistency structure, tensor fusion and refinement encoding are performed on the micro-motion high-frequency response texture map and the skin surface reflection dynamic change image sequence to construct the micro-motion tension tensor.

2. The method for monitoring suspicious persons for bank security as described in claim 1, characterized in that, Constructing a behavioral inertia baseline map, including: Based on each standard behavior trajectory in the multi-path standard behavior trajectory set, discrete-time structured coding is performed to form a trajectory temporal structure unit, which is then used as the temporal structure component of the behavior inertial reference map. Spatial path topology normalization is performed on the standard behavior trajectory set to form trajectory spatial topology structure units, and the trajectory spatial topology structure units are used as spatial structure components of the inertial reference map. Joint modeling of directional stability is performed on the trajectory temporal structure unit and the trajectory spatial topology structure unit to generate a directional stability constraint structure, and the directional stability constraint structure is solidified into the directional stability constraint components of the behavioral inertial reference map; The time structure component, spatial structure component, and directional stability constraint component are fused and encoded in three dimensions to construct a behavioral inertial reference map with fixed inter-layer mapping rules.

3. The suspicious person monitoring method for bank security as described in claim 1, characterized in that, Construct the perturbation field distribution of the target personnel's behavioral trajectory, and generate a behavioral inertial damage distribution matrix based on the difference mapping relationship with the behavioral inertia baseline map, including: A cross-temporal optical flow consistency constraint kernel is introduced into the three-dimensional behavioral image data volume to construct an optical flow constraint consistency tensor field, and the optical flow constraint consistency tensor field is used as the basic structural layer of the behavioral trajectory perturbation field; The visual structural stability distribution features of consecutive image frames are analyzed to construct a local structural stability weight field, which is then used as the stability modulation layer of the behavioral trajectory perturbation field. The optical flow constraint consistency tensor field and the local structural stability weight field are coupled and superimposed to construct a behavior trajectory perturbation field distribution with propagation rules. The distribution of the behavioral trajectory perturbation field and the behavioral inertia reference map are encoded by scale difference mapping to construct a behavioral inertia destruction degree distribution matrix containing spatial hierarchical diffusion characteristics.

4. The method for monitoring suspicious persons for bank security as described in claim 3, characterized in that, The distribution of the behavioral trajectory perturbation field and the behavioral inertial reference map are encoded using scale difference mapping to construct a behavioral inertial destructive degree distribution matrix containing spatial hierarchical diffusion characteristics, including: Based on the interlayer mapping rules of the behavioral inertial reference map, a multi-scale grid alignment structure is constructed to map the behavioral trajectory perturbation field distribution to the corresponding scale space of the multi-scale grid alignment structure. In each scale space, a propagation memory weight structure is introduced to encode the historical perturbation residue superposition of the behavior trajectory perturbation field distribution, thereby generating a scale perturbation accumulation structure. Based on the structural difference relationship between the scale perturbation accumulation structure and the behavioral inertial reference map, a scale difference coupling matrix is ​​constructed; The scale difference coupling matrix is ​​spatially hierarchical diffusion encoded to construct the behavioral inertia destruction degree distribution matrix.

5. The method for monitoring suspicious persons for bank security as described in claim 1, characterized in that, The generation of suspicious risk evolution curves also includes: Based on the aforementioned suspicious risk evolution curve, anomaly warning trigger analysis is performed to construct a warning response; Based on the aforementioned early warning response, early warning dispatch management is implemented, which includes sending early warning information upwards and providing real-time early warnings.

6. The method for monitoring suspicious persons for bank security as described in claim 5, characterized in that, The trigger analysis for abnormal early warning based on the aforementioned suspicious risk evolution curve also includes: If the reporting threshold for the early warning response is not reached, a monitoring instruction is established based on the aforementioned suspicious risk evolution curve. The multispectral imaging device is controlled according to the monitoring instructions to perform monitoring and management of the target personnel.

7. The method for monitoring suspicious persons for bank security as described in claim 1, characterized in that, The generation of suspicious risk evolution curves also includes: Additional data collection is performed on the target personnel, including voice data and network recognition data; Joint authentication of the suspicious risk evolution curve is performed based on the additional data, and monitoring and management are carried out based on the joint authentication results.

8. A suspicious person monitoring system for bank security, characterized in that, The system is used to implement the suspicious person monitoring method for bank security according to any one of claims 1-7, the system comprising: The personnel activity image data stream acquisition module is used to acquire continuous temporal personnel activity image data streams based on multispectral imaging devices deployed in the bank's business area, and to construct a three-dimensional behavioral image data volume including time dimension, spatial dimension and human skeleton posture dimension; The behavioral inertia baseline map construction module is used to read the structural parameters of the bank's business space, the spatial layout parameters of business equipment, and historical normal business samples to build a virtual behavioral simulation environment, generate a set of multi-path standard behavioral trajectories, and construct a behavioral inertia baseline map. The differential mapping module is used to perform cross-temporal optical flow consistency constraint analysis and local visual structure stability calculation of the 3D behavioral image data volume, construct the perturbation field distribution of the target person's behavioral trajectory, and generate the behavioral inertial damage distribution matrix based on the differential mapping relationship with the behavioral inertial reference map. The micro-motion tension tensor construction module is used to perform micro-scale visual enhancement processing on the three-dimensional behavioral image data volume after performing regional localization of the target person. It constructs the micro-motion tension tensor by using sub-pixel level jitter evolution detection of key human skeleton points and dynamic fluctuation analysis of skin surface reflection. The visual risk time-series evolution module is used to input the micro-motion tension tensor and the behavioral inertia destruction degree distribution matrix into the visual risk time-series evolution channel to generate a suspected risk evolution curve.

Images