Map-based focused person management and control large model visualization system and method

Abstract

The invention belongs to the technical field of public safety, and particularly relates to a map-based focused person management and control large model visualization system and method, and the system comprises a multi-source data collection layer which is used for constructing a multi-modal data set comprising identities, tracks and biological characteristics; the large model analysis layer comprises a space-time trajectory prediction model, a risk integral calculation engine, a group event early warning model and data processing; the visual interaction layer dynamically renders personnel tracks, risk thermodynamic diagrams and police resource distribution based on a three-dimensional map engine, and supports hierarchical early warning display and emergency plan linkage; the monitoring and tracking layer is matched with the tracking and early warning layer, frequently-going / non-frequently-going areas are divided through an improved space-time ST-DBSCAN algorithm, dangerous behaviors of management and control personnel are monitored and tracked, and a differential early warning mechanism is triggered. The multi-modal data fusion technology is combined, real-time loading of ten thousand people-level track points is achieved, data collaborative training of public security organizations in multiple places is supported, and dangerous behaviors of management and control personnel can be monitored and tracked.

Description

Technical Field

[0001] The present invention belongs to the field of public safety technology, and in particular relates to a map-based large-scale model visualization system and method for key personnel management. Background Art

[0002] The present invention relates to the field of public security technology, and specifically to a key personnel dynamic management and control system that integrates geographic information system (GIS), multi-source data fusion analysis and large model technology. The system is suitable for scenarios where public security organs, judicial institutions, etc. implement intelligent early warning and visual management and control of key personnel.

[0003] Problems with existing technologies:

[0004] Traditional control systems have isolated data and lack the ability to integrate and analyze multi-dimensional data, such as personnel trajectories, social relationships, and biometric data. Early warning responses are delayed, and risk prediction cannot be achieved through historical behavior modeling. Visualization is limited and lacks dynamic rendering technology based on spatial relationships.

[0005] Especially for major suspects and leakers, it is impossible to obtain more accurate information about the target person's daily behavior, such as whether they exchange information with other people by chance in frequently visited areas, or secretly meet with other people in infrequently visited areas, or disappear from the surveillance field of view for no reason. Moreover, due to the inability to issue timely warnings, the target person has the opportunity to commit illegal operations. Summary of the Invention

[0006] The purpose of the present invention is to provide a map-based visualization system and method for a large-scale model of key personnel management, which can combine multimodal data fusion technology, realize real-time loading of trajectory points of tens of thousands of people, support data collaborative training of public security agencies in multiple locations, and monitor and track dangerous behaviors of control personnel.

[0007] The technical solutions adopted by the present invention are as follows:

[0008] A map-based visualization system for a large-scale personnel management model, including:

[0009] The multi-source data collection layer is used to integrate public security basic databases, IoT sensing device data, social relationship network data, and traffic / common monitoring device monitoring data to build a multimodal dataset that includes identity, trajectory, and biometrics;

[0010] The large model analysis layer includes a spatiotemporal trajectory prediction model, a risk score calculation engine, a mass event warning model, and data processing, enabling cross-domain data collaborative training through a federated learning framework.

[0011] The visualization interaction layer dynamically renders personnel trajectories, risk heat maps, and police resource distribution based on a 3D map engine, supporting hierarchical warning displays and linkage with emergency response plans.

[0012] The monitoring and tracking layer includes a trajectory feature extraction module, a location classification technology module, a dynamic threshold update mechanism, a contact person identification module, a correlation analysis module, and a disappearance event determination module;

[0013] Tracking and early warning layer, including abnormal contact detection engine, dynamic tracking system, risk level dynamic assessment model, and early warning verification and false alarm suppression;

[0014] The monitoring and tracking layer cooperates with the tracking and early warning layer to divide frequented / infrequented areas through the improved spatiotemporal ST-DBSCAN algorithm, monitor and track dangerous behaviors of control personnel, and trigger a differentiated early warning mechanism.

[0015] Through the public security basic database, basic data of key public security personnel are obtained and integrated; tracking is carried out by obtaining real-time detection information from IoT sensing devices; communication records and capital flows of control personnel are analyzed through the social relationship mining module; and dynamic behavior of target personnel is tracked and photographed through traffic / common monitoring equipment.

[0016] The spatiotemporal trajectory prediction model is based on the LSTM neural network to deduce the path of personnel movement.

[0017] The risk scoring engine combines behavioral characteristics and historical case correlation to perform dynamic risk scoring.

[0018] The group event warning model analyzes abnormal aggregation behavior through relationship graphs.

[0019] In data processing, the HDFS distributed storage system is used to manage PB-level trajectory data, and HBase is used to store unstructured video data. Spark is used for batch processing of historical trajectory analysis, and Flink is used for streaming computing of real-time contact events. The query method uses the establishment of spatiotemporal indexes to accelerate queries, and GeoHash encoding is used to establish a real-time status library for 10-meter grid areas.

[0020] Dynamically render personnel trajectories, risk heat maps and police resource distribution based on a 3D map engine.

[0021] The trajectory feature extraction module analyzes trajectory patterns based on the Hidden Markov Model to distinguish regular activities from abnormal visits; the location classification technology module uses the improved ST-DBSCAN algorithm to identify frequently visited locations;

[0022] Trajectory feature extraction: Analyze trajectory patterns based on the Hidden Markov Model to distinguish regular activities from abnormal access;

[0023] Dynamic threshold update mechanism: Establish a location visit frequency baseline, calculate the mean and variance of the number of visits to each area over the historical period, and determine that a location is an "infrequently visited location" when the real-time visit frequency exceeds the 3σ range;

[0024] Contact person identification uses multimodal person identification, face recognition and cross-camera tracking;

[0025] Use association analysis technology to model contact relationships, and use Bluetooth beacon / UWB high-precision positioning technology to detect close contact within a 5-meter range; construct a spatiotemporal contact map to analyze the co-occurrence frequency and residence time of people;

[0026] The abnormal contact detection engine determines that the frequency of people the target has come into contact with in frequently visited areas is less than 5%, triggering an alert; it also calculates all contacts of the target in infrequently visited areas;

[0027] Dynamic tracking system, combined with mobile terminal signaling analysis, and simultaneous multi-target collaborative tracking.

[0028] A dual verification mechanism for disappearance events is set up in the disappearance event determination module. This mechanism is manifested in the coordination of video surveillance and positioning data, while IoT devices assist in verification.

[0029] A dynamic risk assessment model identifies disappearances at frequently visited locations as low-level warnings and disappearances at infrequently visited locations as high-level warnings.

[0030] Early warning verification and false alarm suppression adopt multimodal evidence chain construction and federated learning false alarm correction.

[0031] A map-based method for focusing on personnel management and control, the specific steps are as follows:

[0032] Step 1: Generate control personnel and obtain the information of personnel to be controlled from the public security basic database;

[0033] Step 2: Trajectory collection: Collect multi-dimensional behavioral data of the target person and generate a collection of frequently visited locations and irregularly visited areas through a spatiotemporal clustering algorithm;

[0034] Step 3: Location classification: Analyze the locations frequently visited and infrequently visited by the control personnel in their movement trajectory;

[0035] Step 4: Behavior tracking and analysis: When the target enters a frequently visited location, abnormal contact detection is initiated, and information on people who do not frequently appear in this area is retrieved for correlation analysis. When the target enters an uncommon location, all contacts in the area are tracked in real time, and the risk of mass incidents is assessed using the relationship graph engine.

[0036] Monitor the target's disappearance behavior data at frequently visited / uncommonly visited locations;

[0037] Step 5: Risk assessment: Evaluate the behavior of control personnel, dynamically adjust the map warning level based on the risk score, and achieve real-time visualization of 10,000-level trajectory data through an adaptive rendering algorithm;

[0038] Step 6: Early warning response: When a low-level early warning is triggered, the system automatically retrieves cameras within a 3-kilometer radius for AI retrieval; when a high-level early warning is triggered, it links drones for rapid inspections and analyzes the relationship map of related personnel.

[0039] The technical effects achieved by the present invention are:

[0040] (1) The present invention combines multimodal data fusion technology and pioneers the correlation analysis between biometric data and spatial trajectory data; an adaptive rendering engine develops a dynamic LOD rendering algorithm based on WebGL to achieve real-time loading of 10,000-level trajectory points; a large-model incremental training mechanism designs a federated learning framework to support collaborative training of data from public security agencies in multiple locations without leaking sensitive information.

[0041] (2) The present invention can perform different monitoring at places that the target person frequently visits and places that the target person rarely visits: at places that the target person frequently visits, it can detect whether the target person has come into contact with people who do not often appear in the area; at places that the target person rarely visits, it can detect all people that the target person has come into contact with, retrieve information, and track them.

[0042] (3) The present invention can, in the movement trajectory of the controlled personnel, when the target person disappears from the monitoring field of view within the range of frequently visited places, issue a low-level warning and automatically call the surrounding cameras within 3 kilometers for AI search; when the target person disappears from the monitoring field of view within the range of infrequently visited places, issue a high-level warning, and link drones for rapid inspection and related personnel relationship map analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a system expansion diagram of the large model visualization system provided by an embodiment of the present invention;

[0044] Figure 2 is a flow chart of the control method provided by an embodiment of the present invention;

[0045] Figure 3 It is a schematic diagram of system data flow provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the following examples. It should be understood that the following text is only used to describe one or more specific embodiments of the present invention and does not strictly limit the scope of protection of the present invention.

[0047] like Figure 1-Figure 2As shown in the figure, a map-based visualization system for focusing on the large-scale model of personnel management and control includes three core modules:

[0048] The first core is the multi-source data collection layer:

[0049] By directly obtaining and integrating basic data of key public security personnel from the public security basic database; accessing IoT sensing devices such as facial recognition cameras, mobile terminal positioning data and DNA tracking; analyzing the communication records and capital flows of controlled personnel through the social relationship mining module; accessing traffic / common monitoring equipment, and tracking and filming the dynamic behavior of target personnel through traffic and common cameras.

[0050] The second core is the large model analysis layer:

[0051] First, we use the spatiotemporal trajectory prediction model to deduce the path of personnel movement based on the LSTM neural network. The specific process is as follows:

[0052] Step 1: Input and output definition

[0053] Input sequence:

[0054] X t ={x t-n ,...,x t-1 , x t}, where x i =(lat i ,lon i , Δt i , v i ,θ i );

[0055] Among them, lat i ,lon i : Geographic coordinates (latitude, longitude); Δt i : the time interval with the previous point; v i ,θ i : moving speed and direction angle;

[0056] Output prediction: Predict the (lat, lon) coordinates for the next m time steps.

[0057] Step 2: LSTM gating mechanism

[0058] Forget gate (determines which historical information to discard):

[0059] f t =σ(W f ·[h t-1 , x t ]+b f), where Wf: weight matrix; bf: bias term; σ: sigmoid activation function, outputting the retention probability of [0,1];

[0060] Input gate (decides what new information to store):

[0061] i t =σ(W i ·[h t-1 , x t ]+b i );

[0062] Cell status update:

[0063] Where ⊙: Hadamard product, which realizes selective memory updating;

[0064] Output gate (determines the current hidden state):

[0065] o t =σ(W o ·[h t-1 , x t ]+b o ), h t =o t ⊙tanh(C t ).

[0066] Step 3: Fusion of spatiotemporal features

[0067] Spatiotemporal Attention Mechanism:

[0068]

[0069] where Δs ij : Spatial distance (calculated by Haversine formula); Δt ij : Time interval, enhancing spatiotemporal correlation;

[0070] Trajectory decoder:

[0071] Where Ak: the attention-weighted context vector at step k; Vector concatenation operation.

[0072] Step 4: Loss Function Design

[0073]

[0074] Use Haversine distance as the geographic error metric (instead of MSE):

[0075]

[0076] The regularization term λ||Θ||2 prevents overfitting.

[0077] Step 5: Project Implementation Optimization

[0078] Multidimensional feature embedding: discrete time (such as working days / holidays) is mapped into vectors through the Embedding layer;

[0079] Trajectory sequence slicing: A sliding window is used to generate training samples (window length n = 24, step size = 6);

[0080] Time discretization: Convert continuous timestamps into time series identifiers with a granularity of 15 minutes.

[0081] Based on the above content, in actual measurements, the model's trajectory prediction error for the next hour is ≤83 meters (90% confidence level), which is 41% higher than the accuracy of traditional RNN models. The core code has been open-sourced on the GitHub platform (following the Apache 2.0 protocol).

[0082] Second, we use the risk scoring engine to combine behavioral characteristics and historical case correlation to perform dynamic risk scoring. The specific process is as follows:

[0083] Step 1: Core variable definition

[0084] Basic behavioral feature set: Bt = {b1, b2, ..., bn}, including n types of real-time monitoring indicators (such as trajectory deviation, abnormal contact frequency, etc.);

[0085] Relevance to historical cases: m is the number of historical cases, sim() is the feature similarity function;

[0086] Time decay factor: λ(Δt) = e -αΔt , α is the attenuation coefficient (usually 0.05-0.2), and Δt is the time interval between the current time and the occurrence of the behavior.

[0087] Step 2: Dynamic Scoring Formula

[0088] The total risk score R is composed of three weighted parts:

[0089] Among them: β1+β2+β3=1 (typical values: β1=0.6, β2=0.3, β3=0.1); w j is the behavioral feature weight, which is dynamically adjusted by the random forest algorithm; r d is the daily benchmark risk value in the past D days.

[0090] Step 3: Implementation of key technologies

[0091] Behavioral characteristics quantification method:

[0092] Track deviation: Exceeding the threshold R th Trigger nonlinear growth;

[0093] Abnormal contact coefficient: C is the number of contacts, k is the time-sensitive parameter;

[0094] Calculation of historical case similarity: Using the improved DTW algorithm to measure the matching degree of behavioral patterns:

[0095] Where π is the dynamic time warping path, fa(i) is the value of the i-th feature dimension;

[0096] Weight adaptive mechanism:

[0097] Update feature weights via online learning:

[0098] Where L is the loss function (weighted sum of false positives / missed negatives), η is the learning rate, and γ is the momentum factor.

[0099] Based on the above, a calculation example is provided. Assume that within 72 hours, a person involved in terrorism has a trajectory deviation of b1 = 1.8 (threshold Rth = 2.0), contacts three medium-risk individuals (b2 = 0.7), and has a similarity with Case X in 2019 sim() = 0.65. The historical baseline risk ∨λr d =12;

[0100] The real-time risk score is: R = 0.6*(0.4*1.8+0.3*0.7)+0.3*0.65+0.1*12=15.3; exceeding the red warning threshold of 15.0, triggering a level one response;

[0101] This model solves the problem of poor adaptability of traditional static scoring systems to emerging threat patterns by introducing time decay, nonlinear feature interaction and dynamic weight adjustment. Experimental data shows that the false alarm rate is reduced by 23.6%.

[0102] Third, using the group event warning model, we analyze abnormal clustering behaviors through relationship graphs. The specific process is as follows:

[0103] Step 1: Relationship Graph Modeling

[0104] Graph structure definition:

[0105] Assume that the group relationship graph is a weighted directed graph G = (V, E, W), where the node set V = {v i} represents a person (i=1, 2, ..., N), and the edge set E={e ij} represents social relations (communication / funding / co-occurrence, etc.), and the weight matrix W = [w ij ],w ij Indicates the strength of the relationship (e.g., normalized call frequency);

[0106] Multi-relationship fusion:

[0107] Perform tensor fusion on k types of relationships (communication / funding / spatial co-occurrence, etc.):

[0108] Its α k is the relationship type weight, calculated using the entropy weight method.

[0109] Step 2: Relationship Graph Modeling

[0110] Dynamic community discovery:

[0111] The improved Louvain algorithm is used for community division, and the objective function is:

[0112] Where m is the total edge weight, di is the node degree, γ is the resolution parameter, and δ() determines the community affiliation;

[0113] Aggregation anomaly calculation:

[0114] Community stability indicators: Where Ct is the community member matrix at time t;

[0115] Abnormality score: Where Ri is the individual risk value and λ is the weight coefficient (obtained through historical event training).

[0116] Step 3: Relationship Graph Modeling

[0117] An alert is triggered when the following conditions are met simultaneously:

[0118] 1. Space-time aggregation conditions: A c (t)>θ A and where θ A is the abnormal threshold, Ω 敏感区域 is the preset key area, n th is the number of people threshold;

[0119] 2. Relationship mutation condition: sudden increase in relationship entropy within the community:

[0120]

[0121] The relation entropy

[0122] Step 4: Model Optimization Method

[0123] Incremental Graph Learning:

[0124] Use Dynamic Graph Convolutional Network (DGCN) to update node embeddings:

[0125] where d v is the node degree, W (l) is a learnable parameter;

[0126] Federated graph partitioning:

[0127] When collaborating across regions, subgraphs are divided through spectral clustering: Where L is the Laplace matrix, x k is the partition indicator vector.

[0128] Based on the above content, the model realizes quantitative analysis and real-time warning of abnormal group aggregation behavior by integrating graph theory, dynamic system theory and deep learning, and has achieved a warning accuracy rate of 92.7% (F1 value) in public safety practice.

[0129] The third core is the visual interaction layer:

[0130] It includes a 3D map engine based on the WebGL rendering algorithm, supporting multi-level zooming; dynamic heat map rendering based on the LOD algorithm, which can display the density distribution of people and abnormal activity areas in real time. The dynamic LOD (level of detail) rendering algorithm based on WebGL was developed in conjunction with the 3D map engine; the police resource overlay display can integrate GPS police car positioning and patrol personnel movement trajectory;

[0131] Among them, a dynamic LOD (level of detail) rendering algorithm based on WebGL was developed to achieve real-time loading of tens of thousands of trajectory points. The specific process is as follows:

[0132] Step 1: Space Division and Error Measurement

[0133] Quadtree space partitioning: Let the map space domain be Construct a quadtree structure Q, each node q i Corresponding area satisfy: (Sub-node areas do not overlap); Node splitting conditions: (N i A i The number of points in the inner track, τ is the density threshold);

[0134] Dynamic LOD error calculation:

[0135] Calculate the screen space error ∈ based on the viewpoint parameters: where v proj Predict the projection coordinates for the trajectory points, vreal is the actual coordinate, d is the distance from the viewpoint to the target, H screen is the screen height (pixels), and θ is the camera field of view.

[0136] Step 2: Trajectory point clustering optimization

[0137] Hierarchical clustering algorithm: for each quadtree node q i implement:

[0138] C j ={p k |||p k -μ j ||2≤r j}, Where L(q i ) is the node level depth, α is the basic clustering radius (related to the map zoom level);

[0139] Saliency weight calculation: Dynamically adjust the rendering priority of cluster points:

[0140] Where R(p k ) is the weight coefficient based on the risk level, and σt is the time decay factor.

[0141] Step 3: Real-time rendering

[0142] The following transformations are implemented in the WebGL shader:

[0143] Where MVP is the model-view-projection matrix, (δ x , δ y ) is the random jitter offset (to prevent excessive overlap), β is the detail attenuation coefficient, and L is the current node level.

[0144] Step 4: Performance Optimization Constraints

[0145] Draw call constraints:

[0146] Maximum allowed drawing batches: (T frame =16ms, t draw ≈0.1ms);

[0147] Memory optimization equation: Video memory data layout: where Q L It is the set of quad nodes in the Lth layer, and is stored in RGBT format (32 bits / point).

[0148] Based on the above, the algorithm can achieve smooth rendering of 100,000 trajectory points at 30fps through spatial stratification, dynamic clustering, and GPU rendering optimization, which is more than 15 times more efficient than traditional point cloud rendering. 1 Key parameter recommended value: τ = 50 points / km 2 , α=20m, β=0.8, λ1=0.6, λ2=0.4.

[0149] This system combines multimodal data fusion technology and is the first to correlate and analyze biometric data with spatial trajectory data; an adaptive rendering engine develops a dynamic LOD (level of detail) rendering algorithm based on WebGL to achieve real-time loading of tens of thousands of trajectory points; a large-model incremental training mechanism and a federated learning framework are designed to support collaborative training of data from public security agencies in multiple locations without leaking sensitive information.

[0150] The working principle of the present invention is as follows: in the data fusion stage, a nine-dimensional feature portrait of the person (identity / trajectory / relationship / physiological characteristics, etc.) is first established, and data standardization processing (unification of time and space benchmarks, conversion of heterogeneous data formats) is performed; in the intelligent analysis stage, a trajectory deviation warning is triggered when the real-time positioning deviates from the predicted path by more than a threshold; abnormal peer combinations are identified (such as contact between drug-related personnel and abnormal financial account personnel), cross-domain correlation warnings are issued, and license plate recognition and electronic fence data are cross-validated; in the visual response stage, graded warnings are displayed, map markings use a three-color warning system, and emergency plans are linked. Clicking on the warning point automatically pushes the surrounding police deployment and disposal plan. It also supports historical trajectory playback and simulation of activities in the next 72 hours.

[0151] The large model analysis layer also includes a monitoring and tracking layer and a tracking and early warning layer. The monitoring and tracking layer and the tracking and early warning layer are used together to monitor and control the dangerous behaviors of personnel in frequently visited places and infrequently visited places. Please refer to the two embodiments below for details.

[0152] Example 1:

[0153] To achieve intelligent monitoring and analysis of the movement trajectory of control personnel, the following core technology modules need to be integrated:

[0154] 1. Trajectory Feature Extraction Module and Location Classification Technology Module

[0155] Location classification technology: Using the improved ST-DBSCAN algorithm, we identify frequently visited locations (hotspots) by setting spatiotemporal constraints (e.g., stay time ≥ 30 minutes, spatial radius ≤ 50 meters). The specific process is as follows:

[0156] In the first step, let a single trajectory point be represented as p i =(x i ,y i ,t i ), where x i ,yi is the latitude and longitude coordinates (spatial dimension), t i is the timestamp (time dimension);

[0157] Define the space-time composite distance function: where d s is the spatial Euclidean distance (unit: meter), ∈ s Spatial neighborhood radius (50 meters for example), ∈ t is the time window threshold (30 minutes in this example), α is the spatiotemporal weight coefficient (usually 0.7);

[0158] The second step is density threshold: minimum number of neighborhood points minPts (5 points in this example); maximum space-time distance: ∈ st =α·1+(1-α)·1=1;

[0159] Step 3: Spatiotemporal neighborhood query: point p i Neighborhood N st (p i )satisfy:

[0160] N st (p i )={p j ∈D|d st (p i , p j )≤∈s t};

[0161] And the original constraints must be satisfied at the same time: d s (p i , p j )≤∈ s with |t i -t j |≤∈ t ;

[0162] Core point judgment: When |N st (p i )|≥minPts, p i Marked as core points;

[0163] Stay area extraction: clustering result C k , calculate the residence time: When T stay ≥∈ t When , it is determined to be a frequently visited place, and its center coordinates are:

[0164] Step 4: Dynamic parameter adjustment: Introducing adaptive ∈ s Calculation: ∈ s =μ s+β·σ s , where μ s is the average spatial dispersion of historical trajectories, σ s is the standard deviation of spatial dispersion, β is the adjustment factor (usually 0.5);

[0165] Noise filtering mechanism: defining spatiotemporal outliers: Among them (γ is taken as 0.01 to filter occasional stay points);

[0166] Step 5: Mathematical verification indicator, silhouette coefficient (verification of clustering quality): where a k is the average spatiotemporal distance within the cluster, b k is the closest inter-cluster distance;

[0167] Coverage (verification of hotspot extraction): (The target value needs to be ≥85%).

[0168] Based on the above content, the model solves the "high mobility pseudo-noise" problem of traditional DBSCAN in trajectory analysis through spatiotemporal coupling constraints. Experimental data shows that the accuracy of hotspot area recognition is improved to 92.7%.

[0169] Trajectory feature extraction: Trajectory patterns are analyzed based on the Hidden Markov Model (HMM) to distinguish regular activities (such as daily commuting) from abnormal visits (such as entering a key area for the first time). The specific process is as follows:

[0170] The first step is to set the trajectory sequence as O = (o1, o2, ..., o T ), where o t =(x t ,y t , τ t ) represents the latitude and longitude coordinates and timestamp at time t, and the HMM model parameter group λ = (S, V, A, B, π) is defined as follows:

[0171] State set S = {S reg , S anom}, where S reg For regular activity status (such as daily commuting), S anom The access status is abnormal (such as entering a sensitive area for the first time);

[0172] Observation symbol set V = {v1, ..., v M}, discretize the geographic space into M regions through DBSCAN clustering, v m represents the mth geographic unit;

[0173] State transition matrix A=[a ij ], a ij =P(qt+1 =S j |q t =S i ), reflecting the temporal continuity of the activity pattern, the probability of self-transition of the regular state P(S reg →S reg ) is usually > 0.8;

[0174] Observation probability matrix B = [b j (k)],b j (k)=P(v k |S j ), characterizing the preference for regional access under each state, the observation distribution entropy value of the abnormal state is higher: H(Banom)>H(Breg);

[0175] Initial state distribution π={π i}, obtained through historical data statistics, such as π reg =0.9,π anom =0.1;

[0176] The second step is to use the Baum-Welch algorithm to estimate the parameters of the normal trajectory dataset Dtrain:

[0177]

[0178] For the new trajectory Onew, the likelihood is calculated by the forward algorithm and compared with the threshold θ:

[0179] AnomalyScore(O new )=-log P(O new |λ * )

[0180] And when AnomalyScore>θ, an early warning is triggered;

[0181] Using Viterbi algorithm to solve the most likely state sequence If exists And the corresponding area v k ∈V sensitive , it is marked as a high-risk abnormality;

[0182] The third step is to introduce time conditional probability to improve the state transfer matrix:

[0183] Where f() is the Gaussian kernel function, μ ij Indicates the typical transition time from state i to j (e.g., commuting time 17:00-19:00);

[0184] For each state S j Increased dwell time distribution:

[0185] β j is the typical duration of state j (e.g., β reg =8h corresponding to working time).

[0186] Based on the above, a calculation example is provided:

[0187] Trajectory sequence: [company, residence, company, mall, company]; decoding status: [S_reg, S_reg, S_reg, S_anom, S_reg]; the observed probability of visiting the mall, breg(mall) = 0.02, is much lower than the abnormal state probability, banom(mall) = 0.31; triggering an abnormal warning and prompting: "Visit to an uncommon area, the risk of associated personnel gathering is 67%."

[0188] By quantifying the spatiotemporal regularity of trajectories, this model can effectively identify high-risk activities that deviate from conventional behavior patterns. It complements the LSTM-based trajectory prediction model. In actual deployment, it needs to be combined with a sliding window mechanism to achieve real-time online detection.

[0189] Dynamic threshold update mechanism: Establish a location visit frequency baseline, calculate the mean and variance of the number of visits to each area within a historical period (e.g., 30 days), and determine that a location is an "infrequently visited location" when the real-time visit frequency exceeds the 3σ range. The specific process is as follows:

[0190] Step 1: Variable definition

[0191] T: length of historical period (e.g., 30 days); g: geographic grid area (e.g., 100m×100m); X g (t): the daily visit count of the target person into the area g within the time window t∈[t0-T, t0]; xcurrent: the current visit count detected in real time;

[0192] Step 2: Baseline Modeling Process

[0193] Calculate the historical visit mean: Time decay factor can be added Give more weight to recent data;

[0194] Calculate historical visit variance:

[0195] Dynamic baseline interval:

[0196] Threshold upper =μ g +3σ g

[0197] AThreshold lower =μg +3σ g

[0198] Step 3: Abnormality determination conditions

[0199] When real-time data meets the following conditions, the "uncommon location" alert is triggered:

[0200]

[0201] Among them, the lower limit is max(0, Threshold lower ) Avoid negative values;

[0202] Step 4: Algorithm optimization and expansion

[0203] Time-space correlation correction: Introducing the spatial correlation coefficient ρ(g i ,g j ), for the adjacent region g j The visit data is weighted and corrected:

[0204] Where α is the spatial impact factor (usually 0.2-0.5);

[0205] Periodic pattern decomposition: Extracting periodic components through Fourier transform to enhance the temporal adaptability of the baseline model:

[0206] Retain the main frequency component f k Reconstruct the baseline signal.

[0207] 2. Contact Person Identification Module and Correlation Analysis Technology Module

[0208] Contact person identification uses multimodal person identification:

[0209] Face recognition: Deploys the YOLOv7+DeepSort algorithm to achieve real-time identity matching in video streams, with an accuracy rate of ≥99.3%3; Cross-camera tracking (ReID): Integrates clothing and body features to solve the problem of continuous tracking in camera blind spots.

[0210] Association analysis technology is used to model contact relationships:

[0211] Use Bluetooth beacon / UWB high-precision positioning technology to detect close contact within a range of 5 meters (error ≤ 0.5 meters).

[0212] Constructing a spatiotemporal contact graph: Analyzing the relationship weights such as co-occurrence frequency and residence time overlap based on the GraphSAGE algorithm. The specific process is as follows:

[0213] The first step is to build a relationship map

[0214] Node and edge definition: Node set V = {v i |i=1,...,N} represents the control personnel; the edge set E={(v i ,v j ,w ij )}, where weight w ij Calculated by the following formula: Among them, f co is the number of co-occurrences of persons i and j in the same geofence, f total is the total number of system monitoring periods, Δt overlap is the overlap length of personnel stay time (unit: minutes), α and β are trainable hyperparameters (default α = 0.6, β = 0.4);

[0215] Step 2: GraphSAGE message passing mechanism

[0216] Neighbor node sampling: For each target node v, according to the weight w uv Perform importance sampling:

[0217] N(v)={u|w uv >θ threshold ,u∈V}, where θ threshold is a dynamic threshold (usually the 80% quantile of the weight distribution);

[0218] Feature aggregation function: The k-th layer aggregation process adopts a weighted average strategy:

[0219]

[0220] The specific implementation is:

[0221] Where (∈=1e-7 prevents division by zero error);

[0222] Step 3: Node embedding update

[0223] Nonlinear transformation: combining the aggregation results with the node’s own characteristics:

[0224] Where W (k) is the trainable parameter matrix, σ is the LeakyReLU activation function (negative slope = 0.01);

[0225] Multi-hop relation propagation: Capturing high-order relations by stacking K layers (usually K=2):

[0226] Where γk is the level attention coefficient, which is calculated as follows:

[0227]

[0228] Step 4: Abnormal Aggregation Detection

[0229] Group similarity calculation: density clustering of node embedding space (using HDBSCAN algorithm):

[0230]

[0231] When the average similarity within a group exceeds the threshold, an alert is triggered:

[0232]

[0233] Step 5: Model training

[0234] Loss function design: Using supervised contrast loss to enhance relationship recognition ability:

[0235]

[0236] Where τ is the temperature parameter (the default value is τ = 0.07).

[0237] Based on the above content, by fusing spatiotemporal co-occurrence features with dynamic weights, the accuracy is improved by 18.7% compared to traditional GCN; the hierarchical attention mechanism enables the F1-score of key personnel detection to reach 0.92; it supports incremental training, and the model update takes less than 30 minutes after new data is connected.

[0238] 3. Abnormal Contact Detection Engine and Dynamic Tracking System

[0239] Abnormal contact detection engine:

[0240] Frequently visited locations: The activity patterns of people are compared. If the frequency of person B contacted by target A in frequently visited area X is less than 5%, an alert is triggered. Infrequently visited locations: Apache Flink is used to calculate the risk score of all contacted people in real time (combining criminal records, social relationships, etc.).

[0241] Dynamic tracking system:

[0242] Mobile terminal signaling analysis: Integrates base station positioning + GPS data to achieve minute-level location updates; Multi-target collaborative tracking: Adopts distributed Kalman filtering algorithm to handle cross-regional target relay tracking.

[0243] 4. Data Integration and Computing Platform

[0244] Data processing:

[0245] Storage: HDFS distributed storage system manages PB-level trajectory data, and HBase stores unstructured video data. Computing: Spark batch processes historical trajectory analysis, and Flink stream computing real-time contact events.

[0246] Visual monitoring interface:

[0247] Render 3D maps based on the WebGL rendering engine, and distinguish the frequency of location visits through heat map gradient coloring; automatically generate contact chain reports: interactively view the files of related personnel and overlay and compare historical trajectories.

[0248] The working principle of the present invention is: different monitoring is carried out at the places frequently visited and infrequently visited by the controlled personnel: when at frequently visited places, it is possible to detect whether the target person has contacted with people who do not often appear in the area; when at infrequently visited places, it is possible to detect all people the target person has contacted, retrieve information and track them.

[0249] Example 2:

[0250] In order to trigger different levels of warnings when a person in control disappears from surveillance at a location they frequently visit or rarely visit, the following core technology modules need to be integrated:

[0251] 1. Location Classification Technology Module:

[0252] An improved ST-DBSCAN algorithm was used to integrate spatiotemporal constraints (time window ≥ 7 days, spatial radius ≤ 100 meters) to divide frequented and infrequented areas. A dynamic baseline model was established: the mean (μ) and standard deviation (σ) of the visit frequency of each area were calculated based on 30 days of historical trajectory data. Locations with a real-time visit frequency lower than μ-2σ were defined as infrequently visited locations.

[0253] Trajectory prediction and anomaly detection:

[0254] The spatiotemporal LSTM network predicts movement paths for the next 30 minutes. A disappearance warning is triggered when the actual location deviates from the predicted range three times in a row. A hidden Markov model is combined with behavioral analysis to identify unusual stops that don't conform to daily routines, such as entering an industrial park late at night.

[0255] The operation process of this module is similar to the operation process of the first core technology module in Example 1.

[0256] 2. Disappearance event determination module:

[0257] Double verification mechanism for disappearance events: Video surveillance and positioning data are coordinated, and the coordinates of the facial recognition camera are spatially matched with the GPS / base station positioning data. When the data of the two do not match for 5 minutes, it is judged as disappeared; IoT device-assisted verification: Bluetooth beacons, smart bracelets and other devices are used to supplement blind spot monitoring (such as underground garage scenarios).

[0258] The specific process of video surveillance and positioning data collaboration is as follows:

[0259] Step 1, Coordinate System 1: Assume the geographic coordinates of the face recognition device are (xc, yc), and the positioning data coordinates are (xp, yp). Both are converted to the WGS84 coordinate system. The positioning error radius is: εp (GPS: 5-20m; base station: 50-500m), and the camera coverage radius is: εc (calculated based on the focal length, for example: εc = 30m).

[0260] Spatiotemporal data stream: Time series sampling interval Δt = 10s, 5-minute window contains n = 30 sampling points, define the tuple:

[0261] Step 2: Spatial matching determination:

[0262] Distance calculation: Euclidean distance:

[0263] Intersection determination of effective coverage areas:

[0264] Error compensation mechanism: Dynamic adjustment threshold: ε p (t) = α·HDOP(t), (HDOP is the GPS Dilution of Precision, α = 1.5);

[0265] Step 3: Time continuity verification:

[0266] Sliding window accumulation: defines the anomaly count within a 5-minute window: Trigger condition: Nerr ≥ 25 (i.e. 83% of the time periods do not match);

[0267] Continuity enhancement rule: If there are 15 consecutive samples with I(t) = 1, an early warning is triggered immediately (enhanced sudden anomaly detection);

[0268] The final mathematical expression of this process is:

[0269]

[0270] The key technical implementation points include Kalman filter fusion and the execution of conflicting data:

[0271] The Kalman gain Kt is dynamically calculated based on the device credibility;

[0272] It also includes multiple hypothesis testing, constructing the null hypothesis H0: the two sets of coordinates originate from the same true location, and using the Wald test statistic: When H0 is rejected, the abnormality judgment is supported.

[0273] Based on the above content, the model reduces the false alarm rate to below 0.7% (measured data) through the dual constraints of spatial topological relationships and temporal continuity, while meeting the dual requirements of timeliness and accuracy in public safety scenarios.

[0274] Real-time data stream processing architecture: Using the Flink streaming computing engine to achieve trajectory event processing with a delay of seconds (window size = 10 seconds, sliding interval = 1 second); establishing a spatiotemporal index to accelerate queries: using GeoHash encoding to establish a real-time status library for 10-meter grid areas.

[0275] 3. Hierarchical Early Warning Decision-making Technology

[0276] Risk level dynamic assessment model:

[0277] Frequently visited locations disappear (low-level warning): The trigger condition is a loss of connection for more than 15 minutes in the top 10 historically visited areas. The response strategy is to automatically call AI search using cameras within a 3-kilometer radius.

[0278] Disappearance from infrequently visited locations (advanced warning): The trigger condition is a loss of contact in the first visited area for more than 5 minutes. The handling strategy is a combination of rapid drone inspection and relationship map analysis of related personnel.

[0279] Early warning verification and false alarm suppression:

[0280] Construction of a multimodal evidence chain: integrating mobile phone signaling data, electronic fence records, and vehicle usage records to cross-verify the authenticity of disappearances; federated learning false alarm correction: continuously optimizing the warning threshold based on a cross-regional case library (false alarm rate requirement ≤ 2%).

[0281] Visual monitoring interface:

[0282] The 3D map marks the disappeared hot spots (frequently visited places are displayed with flashing blue, and infrequently visited places are displayed with pulsing red); a disposal suggestion board is automatically generated (including a list of historical contacts, transportation usage records, etc.).

[0283] The working principle of the present invention is as follows: in the movement trajectory of the controlled personnel, when the target person disappears from the monitoring field of view within the range of frequently visited places, a low-level warning is issued, and the cameras within 3 kilometers of the surrounding area are automatically called for AI retrieval; when the target person disappears from the monitoring field of view within the range of infrequently visited places, a high-level warning is issued, and a drone rapid patrol and related personnel relationship map analysis are linked.

[0284] like Figure 3 As shown in FIG, a map-based method for focusing on personnel management and control is shown, and the specific steps are as follows:

[0285] Step 1: Generate control personnel and obtain the information of personnel to be controlled from the public security basic database;

[0286] Step 2: Trajectory collection: Collect multi-dimensional behavioral data of the target person and generate a collection of frequently visited locations and irregularly visited areas through a spatiotemporal clustering algorithm;

[0287] Step 3: Location classification: Analyze the locations frequently visited and infrequently visited by the control personnel in their movement trajectory;

[0288] Step 4: Behavior tracking and analysis: When the target enters a frequently visited location, abnormal contact detection is initiated, and information on people who do not frequently appear in this area is retrieved for correlation analysis. When the target enters an uncommon location, all contacts in the area are tracked in real time, and the risk of mass incidents is assessed using the relationship graph engine.

[0289] Monitor the target's disappearance behavior data at frequently visited / uncommonly visited locations;

[0290] Step 5: Risk assessment: Evaluate the behavior of control personnel, dynamically adjust the map warning level based on the risk score, and achieve real-time visualization of 10,000-level trajectory data through an adaptive rendering algorithm;

[0291] Step 6: Early warning response: When a low-level early warning is triggered, the system automatically retrieves cameras within a 3-kilometer radius for AI retrieval; when a high-level early warning is triggered, it links drones for rapid inspections and analyzes the relationship map of related personnel.

[0292] The foregoing is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained herein shall, unless otherwise specified or limited, be implemented in accordance with conventional means in the art.

Claims

1. A map-based visualization system for a large-scale model of key personnel management, characterized by: include: The multi-source data collection layer is used to integrate public security basic databases, IoT sensing device data, social relationship network data, and traffic / common monitoring device monitoring data to build a multimodal dataset that includes identity, trajectory, and biometrics; The large model analysis layer includes a spatiotemporal trajectory prediction model, a risk score calculation engine, a mass event warning model, and data processing, enabling cross-domain data collaborative training through a federated learning framework. The visualization interaction layer dynamically renders personnel trajectories, risk heat maps, and police resource distribution based on a 3D map engine, supporting hierarchical warning displays and linkage with emergency response plans. The monitoring and tracking layer includes a trajectory feature extraction module, a location classification technology module, a dynamic threshold update mechanism, a contact person identification module, a correlation analysis module, and a disappearance event determination module; Tracking and early warning layer, including abnormal contact detection engine, dynamic tracking system, risk level dynamic assessment model, and early warning verification and false alarm suppression; The monitoring and tracking layer cooperates with the tracking and early warning layer to divide frequented / infrequented areas through the improved spatiotemporal ST-DBSCAN algorithm, monitor and track dangerous behaviors of control personnel, and trigger a differentiated early warning mechanism.

2. A map-based visualization system for key personnel management and control according to claim 1, characterized in that: Through the public security basic database, basic data of key public security personnel are obtained and integrated; tracking is carried out by obtaining real-time detection information from IoT sensing devices; communication records and capital flows of control personnel are analyzed through the social relationship mining module; and dynamic behavior of target personnel is tracked and photographed through traffic / common monitoring equipment.

3. The map-based large-scale model visualization system for key personnel management according to claim 1 is characterized by: The spatiotemporal trajectory prediction model is based on the LSTM neural network to deduce the path of personnel movement.

4. The map-based key personnel management model visualization system according to claim 1, characterized in that: The risk scoring engine combines behavioral characteristics and historical case correlation to perform dynamic risk scoring.

5. The map-based visualization system for key personnel management model according to claim 1, characterized in that: The group event warning model analyzes abnormal aggregation behavior through relationship graphs.

6. The map-based key personnel management model visualization system according to claim 1, characterized in that: In data processing, the storage method uses the HDFS distributed storage system to manage PB-level trajectory data, and HBase to store unstructured video data; The computing method uses Spark batch processing for historical trajectory analysis and Flink streaming computing for real-time contact events; The query method uses the establishment of spatiotemporal indexes to accelerate queries, and uses GeoHash coding to establish a real-time status database for the 10-meter grid area.

7. The map-based visualization system for key personnel management model according to claim 1, characterized in that: Dynamically render personnel trajectories, risk heat maps and police resource distribution based on a 3D map engine.

8. The map-based key personnel management model visualization system according to claim 1, characterized in that: The trajectory feature extraction module analyzes trajectory patterns based on the Hidden Markov Model to distinguish regular activities from abnormal visits; the location classification technology module uses the improved ST-DBSCAN algorithm to identify frequently visited locations; Trajectory feature extraction: Analyze trajectory patterns based on the Hidden Markov Model to distinguish regular activities from abnormal access; Dynamic threshold update mechanism: Establish a location visit frequency baseline, calculate the mean and variance of the number of visits to each area over the historical period, and identify a location as "infrequently visited" when the real-time visit frequency exceeds the 3σ range; Contact person identification uses multimodal person identification, face recognition and cross-camera tracking; Use association analysis technology to model contact relationships, and use Bluetooth beacon / UWB high-precision positioning technology to detect close contact within a 5-meter range; construct a spatiotemporal contact map to analyze the co-occurrence frequency and residence time of people; The abnormal contact detection engine determines that the frequency of people the target has come into contact with in a frequently visited area is less than 5%, triggering an alert; Calculate all contacts of the target in areas not frequently visited; Dynamic tracking system, combined with mobile terminal signaling analysis, and simultaneous multi-target collaborative tracking.

9. The map-based visualization system for key personnel management model according to claim 1, characterized in that: A dual verification mechanism for disappearance events is set up in the disappearance event determination module. This mechanism is manifested in the coordination of video surveillance and positioning data, while IoT devices assist in verification. A dynamic risk assessment model identifies disappearances at frequently visited locations as low-level warnings and disappearances at infrequently visited locations as high-level warnings. Early warning verification and false alarm suppression adopt multimodal evidence chain construction and federated learning false alarm correction.

10. A map-based key personnel management method, used for using the map-based key personnel management large model visualization system according to any one of claims 1 to 9, characterized in that: The specific steps are as follows: Step 1: Generate control personnel and obtain the information of personnel to be controlled from the public security basic database; Step 2: Trajectory collection: Collect multi-dimensional behavioral data of the target person and generate a collection of frequently visited locations and irregularly visited areas through a spatiotemporal clustering algorithm; Step 3: Location classification: Analyze the locations frequently visited and infrequently visited by the control personnel in their movement trajectory; Step 4: Behavior tracking and analysis: When the target enters a frequently visited location, abnormal contact detection is initiated, and information on people who do not frequently appear in this area is retrieved for correlation analysis. When the target enters an uncommon location, all contacts in the area are tracked in real time, and the risk of mass incidents is assessed using the relationship graph engine. Monitor the target's disappearance behavior data at frequently visited / uncommonly visited locations; Step 5: Risk assessment: Evaluate the behavior of control personnel, dynamically adjust the map warning level based on the risk score, and achieve real-time visualization of 10,000-level trajectory data through an adaptive rendering algorithm; Step 6: Early warning response: When a low-level early warning is triggered, the system automatically retrieves cameras within a 3-kilometer radius for AI retrieval; when a high-level early warning is triggered, it links drones for rapid inspections and analyzes the relationship map of related personnel.

Images