Police research and judgment method based on behavior characteristic recognition

Abstract

The application discloses a police research and judgment method based on behavior feature recognition, relates to the technical field of intelligent police and behavior modeling, collects dynamic behavior data of a target object, constructs a dynamic behavior portrait atlas containing behavior events, time nodes, space positions and interactive objects, performs time sequence modeling on the atlas, extracts an individual behavior chain feature vector, compares the current behavior chain with a historical behavior chain and a group average model, calculates a deviation degree, performs cross-region universality evaluation on the deviation degree through a transfer learning model, generates early warning information in combination with regional safety rules and pushes the early warning information to a police research and judgment system if the deviation degree exceeds a preset threshold, and the method can realize accurate identification and hierarchical response of abnormal behaviors, has high robustness, strong generalization capability and good actual combat adaptability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent police work and behavior modeling, and particularly relates to a police work research and judgment method based on behavior feature recognition. BACKGROUND

[0002] Traditional police work research and judgment methods mainly rely on manual experience, clue investigation and historical case analysis, and have problems such as low efficiency, low accuracy and strong subjectivity. In recent years, with the rapid development of artificial intelligence, computer vision and big data technology, intelligent police systems based on behavior recognition and pattern analysis have gradually become a research hotspot. By collecting and analyzing the behavior characteristics of individuals or groups, potential abnormal behaviors can be identified in advance, thereby providing intelligent early warning and decision support for public security organs.

[0003] However, existing behavior feature recognition technologies still have the following deficiencies in actual police applications: the modeling of behavior features relies too much on predefined rules and lacks adaptive learning ability; they fail to comprehensively analyze multi-source heterogeneous data such as geographic location, time node and historical records; and they lack a unified behavior feature label system, making it difficult to standardize data processing.

[0004] Therefore, there is an urgent need for a police work research and judgment method that can integrate multi-source data, automatically learn and recognize suspicious patterns based on behavior features, to improve the intelligence and prediction ability of police work. SUMMARY

[0005] The purpose of the present application is to provide a police work research and judgment method based on behavior feature recognition to solve the problems in the background art.

[0006] To achieve the above purpose, the present application provides the following technical solution: a police work research and judgment method based on behavior feature recognition, comprising:

[0007] Collecting dynamic behavior data of target objects through multi-modal sensors, the data including displacement trajectory, behavior posture, stay duration, interaction mode and social association information;

[0008] Based on the behavior data, constructing a dynamic behavior portrait graph for each target object, the graph containing node entities and their semantic edge relationships;

[0009] Time-series modeling of the behavior portrait graph, extracting individual behavior chain feature vectors using graph neural networks combined with time encoding mechanisms;

[0010] Comparing the current behavior chain with the object's historical behavior chain and the group behavior mean model to calculate the deviation degree;

[0011] Cross-region universality evaluation of the behavior deviation degree through a pre-trained transfer learning model to determine whether it is a potential abnormal behavior;

[0012] If the deviation exceeds the set threshold, the combination area safety rule generates a corresponding level of risk warning and pushes it to the police research and judgment system terminal.

[0013] Preferably, constructing the dynamic behavior portrait atlas of each target object comprises:

[0014] The collected behavior data is segmented by time sequence, and an improved behavior pattern labeling network is used to classify the behavior semantics of each segment of data to generate a preliminary behavior label set;

[0015] Taking the behavior event, the time when the behavior occurs, the spatial position and the interactive object as node entities, a multi-semantic edge is established based on the synchronicity, causality and spatial proximity between nodes;

[0016] The graph structure is input into a dynamic graph construction module that integrates position encoding and semantic attention mechanism to generate a dynamic behavior portrait atlas;

[0017] The evolution trend of the behavior pattern is analyzed by using the edge weight change rate in the atlas, and the atypical path in the atlas structure is identified as a potential abnormal clue.

[0018] Preferably, the improved behavior pattern labeling network is a structure that introduces a timing context perception mechanism and multi-scale feature fusion.

[0019] Preferably, the behavior portrait atlas is time-series modeled and individual behavior chain feature vectors are extracted, comprising:

[0020] The dynamic behavior portrait atlas is sliced according to the time dimension to form a behavior atlas sequence;

[0021] A timestamp is assigned to each behavior node, and the time information is embedded in the node features through a position encoding mechanism;

[0022] A graph neural network model that integrates a time attention mechanism is used to time-series model the behavior atlas sequence;

[0023] The node embedding vectors in each time-series graph are aggregated to construct individual behavior chain feature vectors.

[0024] Preferably, calculating the deviation comprises:

[0025] Aligning the current individual behavior chain feature vector with its historical behavior chain vector in a graph structure;

[0026] Using the historical behavior chain to construct an individual behavior flow distribution model to express the behavior evolution trend in the form of a probability density function;

[0027] A group behavior mean model based on regional time window aggregation is constructed, and a dynamic normalization strategy is introduced to match individual scale differences;

[0028] The multi-scale KL divergence and the atlas embedding difference index are combined to calculate the comprehensive deviation degree between the current behavior chain and the historical and group models, and the behavior abnormality degree is evaluated.

[0029] Preferably, the current behavior chain is divided into multiple sub-sections within a time window, and the Kullback-Leibler divergence value between each sub-section and the historical behavior flow distribution is calculated to form a multi-scale KL deviation sequence; at the same time, the global structure difference measure of the behavior chain and the group mean model in the atlas embedding space is extracted; finally, the KL deviation values at each scale and the graph embedding difference value are input into a weighted sum average discriminant function to output a comprehensive deviation score.

[0030] Preferably, the cross-regional universality evaluation includes:

[0031] A cross-domain sample set containing the source region and the target region behavior atlas is constructed, and a domain adaptive behavior evaluation model is trained using an adversarial transfer mechanism, wherein the model includes a shared encoder and a region discriminant subnetwork; in the evaluation stage, the current behavior deviation degree is input into the shared encoder to obtain a domain-independent feature representation, and is re-projected through a target region feature mapping function; the deviation confidence under the target region behavior baseline is calculated by comparing the distribution of the source domain and the target domain deviation samples, which is used to determine whether it is a potential abnormal behavior.

[0032] In the above technical solutions, the technical effects and advantages provided by the present application are:

[0033] 1. The present application provides a police research and judgment method based on behavior feature recognition, which combines multi-modal data acquisition, dynamic atlas modeling, behavior chain time series analysis and cross-regional transfer recognition, realizes accurate modeling and dynamic evolution tracking of individual behavior patterns. By constructing a behavior portrait atlas and introducing an improved behavior pattern labeling network, the present application can effectively capture complex and hidden abnormal behavior paths, and improve the perception ability and judgment depth of the police system in an open scene.

[0034] 2. The present application uses multi-scale KL divergence and atlas embedding difference fusion to evaluate behavior deviation, and realizes cross-regional universal recognition of behavior deviation through a domain adaptive model, which significantly improves the generalization ability of the system in different regions and different populations. Experimental results show that the proposed method is superior to existing technologies in terms of abnormality recognition accuracy, response timeliness and model stability. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0036] Figure 1 The method mind map of the present application. DETAILED DESCRIPTION

[0037] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0038] Embodiment 1, please refer to Figure 1 The police research and judgment method based on behavior feature recognition described in this embodiment includes:

[0039] Collecting dynamic behavior data of the target object through a multi-modal sensor, the data including displacement trajectory, behavior posture, stay duration, interaction mode and social association information;

[0040] Based on the behavior data, constructing a dynamic behavior portrait atlas of each target object, the atlas containing node entities and semantic edge relationships thereof;

[0041] Time sequence modeling is performed on the behavior portrait atlas, and a graph neural network is used in combination with a time coding mechanism to extract an individual behavior chain feature vector;

[0042] Comparing the current behavior chain with the historical behavior chain of the object and a group behavior mean model, and calculating a deviation degree thereof;

[0043] Through a pre-trained transfer learning model, the behavior deviation degree is evaluated for cross-region universality to determine whether it is a potential abnormal behavior;

[0044] If the deviation degree exceeds a set threshold, a corresponding level of risk warning is generated in combination with regional safety rules, and is pushed to a police research and judgment system terminal.

[0045] In this embodiment, a multi-modal perception behavior data acquisition system for a police service scenario is proposed, which is used to realize holographic acquisition of dynamic behavior data of a target object, and to provide support for subsequent behavior portrait atlas construction and behavior chain analysis.

[0046] The system is deployed in key areas of the city, including entrances and exits, squares, subway stations, commercial centers, etc., using a fusion sensor architecture, which specifically includes:

[0047] Positioning sensor module: Deploy high-precision UWB positioning base station and video target re-identification (Re-ID) unit, real-time acquisition of target object displacement trajectory through multi-source position information fusion algorithm, accuracy control within 0.5 meters, trajectory data recorded in the form of timestamp sequence.

[0048] Behavior and posture perception module: Based on distributed camera array and posture recognition model (such as OpenPose or its lightweight variant), extract the skeletal key point sequence and posture stability parameters of the target object at a specific moment, output one frame of action vector per second.

[0049] Distributed camera and posture recognition are used to describe the deployment strategy (such as setting up 1 camera every 50 meters). Use the model (such as OpenPose-Lite) to describe the frame rate and key point recognition accuracy (such as > 85% accuracy). The output feature is a vector composed of 18 skeletal key point coordinates per frame.

[0050] Stay duration calculation module: Combined with trajectory speed vector and regional stagnation threshold, if the target speed is less than 0.2 m / s within 3 meters for more than a set time (such as 30 seconds), it is marked as a "stay event", and the corresponding stay location, duration and place semantic label are recorded.

[0051] Interaction mode recognition module: Based on image semantic analysis and spatio-temporal contact discrimination method, identify the relationship patterns such as walking together, meeting, following, etc. between the target and other individuals, output the interaction label sequence, and construct the preliminary interaction graph.

[0052] Social association reasoning module: Combined with historical interaction frequency, interaction duration and spatial overlap rate, construct the initial social association graph, which is used for subsequent social behavior modeling and group behavior analysis.

[0053] The collected multi-modal data is pre-processed by the edge node and uploaded to the behavior data fusion module of the police research and judgment platform, formatted using a unified data structure (such as JSON nested object), supporting subsequent portrait graph generation and time series modeling tasks.

[0054] The embodiment provides an implementation method for constructing a dynamic behavior portrait graph of a target object, which is suitable for a police research and judgment system based on multi-modal behavior data. The key is to accurately express the individual behavior evolution process through graph modeling, and to carry out abnormal behavior recognition based on graph structure.

[0055] Behavior event segmentation and semantic labeling: The raw behavior data received includes the displacement trajectory of the target object, the posture sequence, the interaction information, and the time label. The system first segments the data according to the time window, and each segment corresponds to a possible behavior segment. Then, each segment is input into the improved behavior pattern labeling network (BPN-Net) for semantic classification. This network introduces a timing context perception mechanism, captures the previous and subsequent behavior states through a bidirectional gated recurrent unit (Bi-GRU), and fuses multi-scale convolution feature extraction channels (such as different kernel width combinations of 1D-CNN) to enhance the discriminability of weak semantic behavior segments. The network outputs a behavior label sequence, which serves as the semantic basis for subsequent graph nodes.

[0056] The structure and implementation method of the improved behavior pattern labeling network (BPN-Net) are as follows:

[0057] Input layer: receives the segmented behavior data segments (such as skeletal posture sequences, trajectory changes).

[0058] Timing modeling: Bi-GRU (bidirectional gated recurrent unit) is used to extract previous and subsequent context information.

[0059] Feature fusion: three parallel 1D convolution paths (e.g., kernel width 3, 5, 7) are set up for multi-scale feature extraction.

[0060] Classification layer: the Softmax output layer determines the behavior categories.

[0061] The training process includes: data set sources (such as urban combat data or simulation scenarios), labeling methods. The loss function uses cross entropy loss. The optimizer uses Adam, and the basic parameters such as learning rate and training rounds are set. The precision evaluation indicators (Precision, Recall) are used for parameter tuning.

[0062] Node entity and semantic edge construction: taking each behavior event as the center node, its behavior occurrence time, spatial position, and involved interaction objects are extracted as the extended attributes of the node entity. The system defines three types of edge relationships:

[0063] Synchronic edge, representing the simultaneous occurrence of multiple behavior events in similar time periods;

[0064] Causal relationship edge, connecting the evolution sequence before and after the behavior based on the event sequence logic and context prediction model;

[0065] Spatial proximity edge, representing the spatial proximity between event occurrence locations. Each edge is attached with a weight, reflecting the relationship strength or frequency.

[0066] Dynamic graph structure generation: the constructed initial graph structure is input into the dynamic graph construction module that integrates location encoding and semantic attention mechanism. This module performs location embedding on event timestamps and identifies important node relationships through graph attention networks; it continuously updates the graph through a sliding window approach to form a dynamically constructed dynamic behavior portrait graph.

[0067] Behavior evolution analysis and abnormal path identification: the rate of change of edge weights over time is monitored on the dynamic graph to calculate the evolution trend of behavior relationships. If a node path significantly deviates from the individual historical model and the group average evolution trajectory in terms of edge weight growth rate, it is marked as an atypical path. This path is used as a potential abnormal clue in the early warning process.

[0068] The dynamic behavior portrait graph constructed by the method described in this embodiment has high plasticity and expressiveness, can support fine-grained behavior recognition and continuous behavior evolution modeling, and significantly improves the intelligent level and early warning accuracy of the police research and judgment system.

[0069] The embodiment provides a behavior portrait graph time series modeling method based on a graph neural network, which is used to extract a behavior chain feature vector of a target object, and realize modeling of a behavior evolution trend and subsequent deviation analysis.

[0070] Behavior graph time slice construction: the generated dynamic behavior portrait graph is sliced in the time dimension, a sliding window size and step (such as 30 minutes / 15 minutes) are set, a series of graph slices are formed on the timeline to constitute a behavior graph sequence. All behavior nodes and their semantic edge relationships in the time period are retained in each slice to ensure the local continuity and context integrity of the graph structure.

[0071] Node time encoding embedding: each behavior node in the graph sequence is assigned its corresponding absolute timestamp. A sine-cosine position encoding function is used to convert the time information into a fixed-dimensional vector and embed it into the original node feature vector, realizing the fusion of time attributes and semantic attributes. This mechanism enhances the model's ability to perceive the sequence and rhythm changes of events.

[0072] Time series graph neural network modeling:

[0073] A graph neural network model integrating time attention mechanism is constructed, which consists of two main parts:

[0074] Structure layer: based on a multi-layer graph attention network (GAT), each graph slice is processed to extract local behavior structure relationships;

[0075] Time fusion layer: a time attention mechanism (such as time-aware Transformer or time-gated unit) is introduced to perform cross-time fusion of node representations in the graph sequence dimension, capturing implicit paths and abnormal transfer points of behavior evolution.

[0076] Behavior chain feature vector generation: Global pooling (such as mean pooling, attention pooling) is performed on the node embedding vectors in each graph slice to obtain the behavior state representation of the time slice. Then, the graph representations of all time slices are spliced and projected to a unified feature space through a nonlinear transformation to form a complete individual behavior chain feature vector. This vector not only reflects the semantic and structural features of the behavior, but also retains its time-varying trend.

[0077] Through this embodiment, the system can accurately capture the evolution context of the target object behavior in a complex behavior graph sequence, provide high-dimensional and stable behavior chain representation for behavior deviation degree calculation and cross-region anomaly recognition, and significantly enhance the behavior understanding and prediction ability of the system.

[0078] This embodiment provides an individual behavior deviation degree calculation method for comprehensively evaluating the abnormality degree of the current behavior chain based on individual historical behavior and group behavior benchmarks. This method combines graph structure alignment, behavior flow modeling, multi-scale KL divergence analysis, and graph embedding difference fusion, and is suitable for precise identification of potential abnormal behavior in complex police research and judgment scenarios.

[0079] Graph structure alignment and behavior vector registration: First, the feature vector of the current behavior chain is aligned with its historical behavior chain in graph structure. Structure-preserving contrastive embedding mechanisms (such as graph isomorphism regularization or topology-preserving mapping) are used to align the node positions, semantic relationships, and evolution order in the two behavior chain graphs, ensuring that similar behavior structures are consistent in the vector space and reducing the deviation caused by behavior feature drift.

[0080] The alignment methods include:

[0081] Node-level alignment: Nodes are matched based on behavior event time and semantic similarity.

[0082] Edge structure alignment: The structural similarity between graphs is calculated using the shortest path distance or Jaccard similarity.

[0083] Individual behavior flow modeling: Based on the historical behavior chain sample set, the occurrence probability of each behavior type in a continuous time period is calculated, and the kernel density estimation is used to generate a probability density function (PDF) model of the individual behavior flow, which is used to express the distribution characteristics of behavior evolution trend and behavior intensity, and constitutes the individual benchmark model.

[0084] Group behavior mean model and normalization mechanism: aggregate multiple individual behavior chain samples in a specified area time window, calculate the graph structure mean embedding, and form a group behavior mean model. Considering the existence of bias in different individual behavior scales, a dynamic normalization strategy is introduced to adjust the current behavior chain feature value based on the individual historical behavior variance to realize the scale alignment with the group mean model.

[0085] The group behavior mean model is used to describe the behavior chain feature vector of multiple individuals collected in the same period and the same area. The group mean is represented using weighted average or K-Means cluster center. The current behavior chain vector of each individual is normalized using the standard deviation of its historical behavior features.

[0086] Multi-scale KL divergence and graph embedding difference analysis: divide the current behavior chain into multiple time sub-sections, calculate the Kullback-Leibler divergence (KL divergence) between the behavior probability distribution of each sub-section and the individual historical PDF model, and construct a multi-scale KL deviation sequence.

[0087] At the same time, the global structure difference between the current behavior chain graph embedding representation and the group mean model in the graph embedding space is measured, such as using graph cosine similarity, graph embedding distance, or manifold contrast index to evaluate the structural difference.

[0088] Finally, all sub-section KL divergence values and graph embedding structure difference indicators are input into a weighted sum average discriminant function. The weighting coefficients can be preset based on the task scenario or learned from the training data. The function outputs the final comprehensive deviation score, which is used to evaluate the abnormality of the current behavior chain under the individual and group benchmarks, and serves as the input of the subsequent warning module in the police system.

[0089] Specifically, the complete behavior chain is divided into N time sub-sections (such as 5 minutes per section), and the probability density function is modeled using kernel density estimation for a behavior feature (such as "stay duration" and "node jump frequency") in the historical behavior chain. The kernel function can use Gaussian kernel, and the bandwidth parameter is determined by Silverman rule. KL divergence is calculated at different scales (such as 5min / 10min / 15min). Finally, the multi-scale values are weighted and averaged (weights can be trained or empirically set) as the comprehensive deviation value.

[0090] The embodiment provides a deviation evaluation mechanism that combines time series modeling and graph structure analysis, taking into account behavior semantics, structural changes, and temporal evolution characteristics, significantly improving the robustness and recognition accuracy of anomaly detection, and being suitable for various police research and judgment needs such as cross-region migration and complex scene monitoring.

[0091] The embodiment provides a method for behavior deviation degree cross-region universality evaluation, which is suitable for police behavior recognition tasks in the presence of regional differences or environment migration. By introducing an adversarial migration mechanism to construct a domain adaptive behavior evaluation model, migration generalization of the target region data after the source region model training is realized, and the model performance degradation problem caused by the regional behavior distribution difference is solved.

[0092] Firstly, behavior graph data of a source region (for example, city A) and a target region (for example, city B) are collected, and behavior chain feature vectors and corresponding deviation labels are extracted, respectively. A cross-domain behavior sample set is constructed, which includes source domain normal behavior chains, source domain abnormal behavior chains, target domain unlabeled behavior chains and the like.

[0093] An adversarial migration structure including a shared encoder and a regional discrimination subnetwork is constructed. The shared encoder is used to extract domain-independent potential feature representation of the behavior chain, and the output thereof serves as a basis for cross-region migration. The regional discrimination subnetwork is used to predict whether the input sample belongs to the source domain or the target domain. In the training stage, the domain difference is maximally confused by using an adversarial loss function (such as a Gradient Reversal Layer), so that the ability of the encoder to extract domain-insensitive features is improved. The prediction subnetwork of the behavior deviation degree is optimized on the source domain sample, and is used to learn the deviation score.

[0094] In the evaluation stage, the behavior chain of the current target object and the deviation degree vector thereof are input into the shared encoder to obtain the domain-independent feature representation thereof. The representation is further input into a target region feature mapping function (such as a target domain projection space constructed by an automatic encoder or a variational reconstruction network), and is re-projected into a target region behavior representation space.

[0095] The deviation sample distribution extracted in the training stage of the source domain and the target domain is used as a behavior baseline, the re-projected features are compared in distribution by using maximum mean difference (MMD), kernel distance or energy score, and the deviation confidence value of the current behavior chain under the target region behavior model is calculated. If the confidence exceeds a preset threshold, the system will mark the behavior as a potential abnormal behavior, and enter a subsequent police warning process.

[0096] By using the adversarial migration and domain alignment technology, the trained model of the source region is efficiently migrated to the target region without retraining the target region model, good scalability and migration stability are achieved, and the behavior abnormality recognition capability of the cross-region police research and judgment system is significantly improved.

[0097] The embodiment provides a police warning generation method based on fusion of deviation degree threshold and regional safety rules, which is used for automatically triggering a police risk warning of a corresponding level in combination with a regional context condition when the individual behavior deviation degree is significant, and pushing to a research and judgment system terminal to realize instant response and auxiliary decision-making.

[0098] The system first receives the individual behavior chain comprehensive deviation score output by the comprehensive deviation analysis module. The score is compared with preset multi-level risk threshold values, which are set as three-level trigger modes (for example: slight abnormality, moderate abnormality, and significant abnormality), and each level corresponds to a different security response level. The threshold values can be dynamically adjusted based on historical police case data and expert annotation models.

[0099] To avoid false positives triggered only by numerical deviations, the system introduces a regional security rule engine that dynamically corrects based on context information such as the location of the behavior (e.g., key places, high-crime areas), time period (night / commuting peak hours), and surrounding risk level.

[0100] For example, if a deviation behavior occurs in a transportation hub area or during a key period of public security control, its risk level will be increased by one level; if the behavior is in a regular personnel gathering area and short-term abnormality occurs frequently, the early warning can be delayed.

[0101] Combining the deviation level and regional rule weight, the system finally determines the early warning level label of the behavior event (such as level I attention, level II patrol, and level III intervention), and constructs a standardized behavior risk event data structure, which includes fields such as target identification, behavior chain ID, deviation feature summary, geographic location, confidence value, and determination rule hit situation.

[0102] The risk event is pushed in real time to the police research and judgment system terminal through a security communication channel, supporting multi-platform response, including the on-duty command platform, patrol police car terminal, and mobile law enforcement APP. The terminal interface pops up an early warning notification, while associating historical behavior trajectory, group portrait, and event heat map to assist police personnel in quickly researching and disposing.

[0103] Based on the implementation of behavior deviation anomaly detection, this embodiment combines spatial, temporal, and regional security elements to dynamically determine the risk level and implement real-time pushing to multiple terminals in a standardized data form, significantly improving the response efficiency, accuracy, and resource allocation capability of the police system.

[0104] Example 2, to verify the effectiveness of the "police research and judgment method based on behavior feature recognition" proposed in the present application in actual application, the applicant builds a multi-region simulation experiment platform, constructs an experimental data set based on the real combat data of a certain city public security, and compares and tests it with existing typical methods (such as traditional RNN behavior recognition model and static rule engine), including core indicators such as abnormal behavior recognition accuracy, response time, and cross-region migration adaptability.

[0105] Data source: Select the key place video monitoring and location trajectory data of two regions of City A and City B for nearly 6 months, containing the behavior records of 1,800 target objects;

[0106] Behavior categories: 12 normal behaviors (e.g. walking, waiting, commuting) and 3 abnormal behaviors (e.g. abnormal loitering, reverse walking, crowd scattering) are covered.

[0107] Annotation method: Three police officers jointly annotate behavior labels, construct a multi-modal behavior graph, and corresponding behavior chain feature set.

[0108] Method comparison:

[0109] Scheme A (the present application): dynamic graph modeling + BPN-Net semantic annotation + cross-domain transfer recognition.

[0110] Scheme B (control group 1): traditional RNN behavior sequence modeling without graph structure.

[0111] Scheme C (control group 2): static rules + abnormal threshold matching method.

[0112] Test scene: Deploy the model trained in city A in the middle of city B to test the cross-region recognition ability.

[0113] Evaluation index: abnormal recognition accuracy (precision), recall rate, F1 score, response time (milliseconds).

[0114] Table 1: Experimental result summary table

[0115] Method Accuracy (%) Recall (%) F1 Score Average response latency (ms) Invention Scheme A 91.6 89.3 0.904 380 Control B (RNN) 83.2 75.4 0.791 620 Control C (Rules) 68.5 63.1 0.656 240

[0116] As shown in Table 1, the accuracy and recall rate of the abnormal behavior recognition of the present application are significantly better than those of the control method, which reflects the advantages of dynamic graph time series modeling and deviation degree fusion recognition mechanism. Although the response time is slightly higher than that of the static rule system, it still maintains a low delay while maintaining high precision recognition, which is suitable for actual police rapid response scenarios. In cross-region deployment, the present application relies on the anti-migration mechanism to achieve high-robustness migration recognition, which can maintain stable recognition performance without re-labeling data in the target area, significantly reducing labor costs.

[0117] The experimental results show that the technical scheme proposed in the present application is superior to the prior art in terms of abnormal behavior recognition accuracy, time efficiency and cross-domain adaptability, has obvious practical value and popularization potential, and achieves the beneficial technical effects claimed in the present application.

[0118] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A police research and judgment method based on behavior feature recognition, characterized in that: The method comprises the following steps: Collecting dynamic behavior data of target objects through multi-modal sensors, the data including displacement trajectory, behavior posture, stay duration, interaction mode and social association information; Based on the behavior data, constructing a dynamic behavior portrait atlas for each target object, the atlas containing node entities and their semantic edge relationships; Performing time-series modeling on the behavior portrait atlas, and extracting individual behavior chain feature vectors using graph neural networks combined with time encoding mechanisms; Comparing the current behavior chain with the object's historical behavior chain and the group behavior mean model to calculate its deviation degree; The calculation of deviation degree includes: aligning the current individual behavior chain feature vector with its historical behavior chain vector in graph structure; constructing an individual behavior flow distribution model using the historical behavior chain, expressing the behavior evolution trend in the form of probability density function; constructing a group behavior mean model based on regional time window aggregation, and introducing a dynamic normalization strategy to match individual scale differences; using multi-scale KL divergence and graph atlas embedding difference indicators to jointly calculate the comprehensive deviation degree between the current behavior chain and the historical and group models, and evaluating the degree of behavior anomaly; Through a pre-trained transfer learning model, the behavior deviation degree is evaluated for cross-regional universality to determine whether it is a potential abnormal behavior; If the deviation degree exceeds the set threshold, generate a corresponding level of risk warning according to the regional safety rules, and push it to the police research and judgment system terminal.

2. The method of claim 1, wherein the method further comprises: The method for constructing a dynamic behavior portrait atlas for each target object comprises the following steps: Segmenting the collected behavior data according to time sequence, and using an improved behavior pattern labeling network to classify each segment of data and generate a preliminary behavior label set; Taking behavior events, time of behavior occurrence, spatial location and interactive objects as node entities, and establishing multi-element semantic edges based on the synchronicity, causality and spatial proximity between nodes; Inputting the graph structure into a dynamic graph construction module with position encoding and semantic attention mechanism to generate a dynamic behavior portrait atlas; Using the edge weight change rate in the graph to analyze the behavior pattern evolution trend and identify atypical paths in the graph structure as potential abnormal clues.

3. The method according to claim 2, wherein the method is characterized in that: The improved behavior pattern labeling network introduces a time context perception mechanism and a multi-scale feature fusion structure.

4. The method of claim 1, wherein the method further comprises: The method for performing time-series modeling on the behavior portrait atlas and extracting individual behavior chain feature vectors comprises the following steps: Slicing the dynamic behavior portrait atlas according to the time dimension to form a behavior atlas sequence; Assigning a timestamp to each behavior node and embedding the time information into the node features through a position encoding mechanism; Using a graph neural network model with time attention mechanism to perform time-series modeling on the behavior atlas sequence; Aggregating node embedding vectors in each time-series graph to construct an individual behavior chain feature vector.

5. The method of claim 1, wherein the method further comprises: Divide the current behavior chain into multiple sub-segments within a time window, and calculate the Kullback-Leibler divergence value between each sub-segment and the historical behavior flow distribution to form a multi-scale KL deviation sequence; at the same time, extract the global structure difference measure of the behavior chain and the group mean model in the graph embedding space; finally, input the KL deviation values and graph embedding difference values of each scale into a weighted sum average discriminant function to output a comprehensive deviation score.

6. The method of claim 1, wherein the method further comprises: The cross-region generality evaluation includes: A cross-domain sample set containing source region and target region behavior atlas is constructed, and a domain adaptive behavior evaluation model is trained by using an adversarial transfer mechanism, wherein the model contains a shared encoder and a region discrimination subnetwork; in the evaluation stage, the current behavior deviation degree is input into the shared encoder to obtain domain-independent feature representation, and is re-projected through a target region feature mapping function; the deviation confidence under the target region behavior baseline is calculated by combining the distribution comparison of the source domain and the target domain deviation samples, and is used to determine whether it is potential abnormal behavior.

Images