Behavior Abnormality State Recognition and Early Warning Method Based on Video Image Processing

Abstract

The present invention provides a method for recognizing and warning of abnormal behavior states based on video image processing, including: collecting basic data, extracting scene-adaptive features, and generating scene-adaptive behavior feature vectors; performing multi-dimensional abnormal behavior energy evaluation and grading on the scene-adaptive behavior feature vectors to obtain an abnormal behavior energy grading matrix; performing dynamic resource optimization and multi-camera collaborative tracking based on the abnormal behavior energy grading matrix to obtain a cross-camera abnormal event association graph; verifying the cross-camera abnormal event association graph and generating hierarchical warning information, and at the same time evaluating the resource allocation efficiency and generating optimized configuration parameters to feedback to the resource optimization step. The present invention improves the accuracy of abnormal behavior recognition and the utilization efficiency of system resources through multi-dimensional abnormal energy evaluation, dynamic resource optimization, and multi-camera collaborative tracking.

Description

Technical Field

[0001] The present invention relates to a behavior recognition method, in particular to a method for recognizing and warning abnormal behavior states based on video image processing. Background Art

[0002] In places such as smart cities, important infrastructure, transportation hubs, and large commercial complexes, real-time detection and warning of potential abnormal behaviors can effectively prevent safety incidents and minimize casualties and property losses to the greatest extent.

[0003] Existing abnormal behavior recognition technologies are mainly divided into two categories: rule-based methods and statistical learning-based methods. Rule-based methods identify abnormalities by predefined a series of behavior patterns and judgment criteria, such as detecting the movement speed with a fixed threshold, judging the trajectory deviation, and detecting the area intrusion. Such methods are simple to implement and have high computational efficiency, but lack adaptability to complex scenarios and the rule setting is too rigid. Statistical learning-based methods construct a statistical model of normal behaviors and identify behaviors that deviate from the normal pattern as abnormal, such as Gaussian mixture models, autoencoders, and one-class support vector machines. Such methods perform well in simple scenarios but lack robustness to changing environments. In addition, most existing systems adopt a fixed sampling rate and a uniform resource allocation strategy, lacking differential processing of the importance of different scenarios and events. Traditional multi-camera systems usually adopt a simple neighborhood information sharing mechanism, and each camera node works independently, lacking an effective cooperation mechanism and resource optimization strategy.

[0004] The core problems faced by current abnormal behavior recognition technologies are the singularity of the abnormal evaluation model and the static nature of resource allocation. In terms of abnormal evaluation, traditional methods often solely rely on spatial distance metrics of video images (such as Mahalanobis distance or Euclidean distance) to evaluate the degree of abnormal behavior, ignoring the changing characteristics in the time dimension and behavior context information, resulting in the system being difficult to distinguish temporary deviations from persistent abnormalities, and also unable to capture the local aggregation effect and evolution trend of abnormal behaviors. Nor does it integrate other information, such as activity information in the form of text, voice, etc.

[0005] This one-dimensional evaluation leads to a high false alarm rate, especially more obvious in crowded or complex environments. In terms of resource allocation, most existing systems adopt a preset and static resource allocation strategy, ignoring the dynamic nature of abnormal events and the need for differential processing. Simply relying on video image processing, the resource occupancy rate is high. When the system faces multiple abnormal events occurring simultaneously, it may cause key abnormal events to be delayed due to insufficient resources, while resulting in inefficient utilization and waste of computing resources. Summary of the Invention

[0006] Objective of the invention: To provide a method for identifying and warning abnormal behavior states based on video image processing. By combining multi-modal information such as activity information, user-set data, and activity information, the speed of video classification and processing is improved, and then abnormal regions are quickly screened out. More resources are then allocated to the abnormal regions to quickly identify abnormal states, with the expectation of solving at least one technical problem existing in the prior art.

[0007] Technical solution: A method for identifying and warning abnormal behavior states based on video image processing includes the following steps:

[0008] Collect basic data and perform scene adaptive feature extraction based on it to generate scene-adaptive behavior feature vectors; the basic data includes original video stream data, user-set data, and activity information data;

[0009] Perform multi-dimensional abnormal behavior energy evaluation and grading on the scene-adaptive behavior feature vectors to obtain an abnormal behavior energy grading matrix;

[0010] Based on the abnormal behavior energy grading matrix, perform dynamic resource optimization and multi-camera collaborative tracking to obtain a cross-camera abnormal event association graph;

[0011] Verify the cross-camera abnormal event association graph and generate hierarchical warning information.

[0012] In another embodiment of the present application, it further includes evaluating the resource allocation efficiency, generating optimized configuration parameters, and feeding them back to the resource optimization step.

[0013] According to one aspect of the present application, the step of performing multi-dimensional abnormal behavior energy evaluation and grading to obtain an abnormal behavior energy grading matrix includes:

[0014] Based on the historical data of the scene-adaptive behavior feature vectors, construct a time-sensitive background model with memory decay characteristics, establish the boundary of normal behavior, and extract a typical abnormal pattern library;

[0015] Based on the typical abnormal pattern library, construct an abnormal energy function; for the current scene-adaptive behavior feature vector, apply the abnormal energy function to calculate the abnormal energy value including spatial distance, time distance, complexity coefficient, and business impact factor, and generate an abnormal behavior index;

[0016] Establish a regional value map, perform multi-dimensional comprehensive evaluation in combination with the abnormal behavior index, grade the abnormal behavior, and generate an abnormal behavior energy grading matrix.

[0017] According to one aspect of the present application, the step of applying the abnormal energy function to calculate the abnormal energy value including spatial distance, time distance, complexity coefficient, and business impact factor and generating an abnormal behavior index includes:

[0018] Construct an abnormal energy function \(E(b)=\alpha\cdot D_s(b,M)+\beta\cdot D_t(b,H)+\gamma\cdot C(b)+\delta\cdot I(b)\);

[0019] Where \(D_s(b,M)\) represents the spatial distance between behavior \(b\) and the normal behavior boundary \(M\), \(D_t(b,H)\) represents the temporal distance between behavior \(b\) and the historical behavior sequence \(H\), \(C(b)\) represents the complexity coefficient of behavior \(b\), \(I(b)\) represents the business impact factor of behavior \(b\), and \(\alpha\), \(\beta\), \(\gamma\), \(\delta\) are adjustable weight parameters;

[0020] Apply the abnormal energy function to the current scene adaptation behavior feature vector to calculate the abnormal energy value;

[0021] Based on the abnormal energy value, combined with the preset judgment criteria, determine the degree of behavior abnormality and generate an abnormal behavior index.

[0022] According to one aspect of the present application, the steps of establishing a regional value map, conducting multi-dimensional comprehensive evaluation in combination with the abnormal behavior index, classifying abnormal behaviors, and generating an abnormal behavior energy classification matrix include:

[0023] Based on the business importance of the preset monitoring area, construct a regional value map representing the relative value of each area;

[0024] Weightedly combine the abnormal behavior index with the regional value map to construct a weighted abnormal importance index reflecting the importance of abnormal behaviors;

[0025] Apply the analytic hierarchy process algorithm to conduct multi-dimensional comprehensive evaluation of the weighted abnormal importance index according to the pre-stored abnormal behavior types, occurrence locations, and influence scopes to obtain an evaluation result;

[0026] Based on the evaluation result, classify the abnormal behaviors into five levels: critical, high-risk, medium-risk, low-risk, and prompt, and generate an abnormal behavior energy classification matrix.

[0027] According to one aspect of the present application, the weight parameters \(\alpha\), \(\beta\), \(\gamma\), \(\delta\) in the abnormal energy function are dynamically adjusted through the following steps:

[0028] Collect historical data containing confirmed abnormal behaviors and extract historical abnormal detection results;

[0029] Based on the historical abnormal detection results, apply the particle swarm optimization algorithm to calculate the weight parameter combination that maximizes the detection accuracy;

[0030] Update the weight parameters \(\alpha\), \(\beta\), \(\gamma\), \(\delta\) in the abnormal energy function, and maintain the constraint condition of \(\alpha+\beta+\gamma+\delta = 1\);

[0031] Periodically adjust the weight parameters according to the changes in the current scene to make the abnormal energy function adapt to different scene characteristics.

[0032] According to one aspect of the present application, after calculating the abnormal energy value, it further includes the step of calculating the local energy density:

[0033] Construct a local energy density function ρ(b, r) to represent the abnormal energy density within a radius r around behavior b; ρ(b, r)=∑E(bi)·w(d(b, bi)), where bi is other behavior within radius r, E(bi) is its abnormal energy value, d(b, bi) is the distance between behavior b and bi, and w is a distance-based weight function;

[0034] Adopt a Gaussian kernel function as the weight function w(d)=exp(-d 2 / 2σ 2 ), where σ is an adjustable influence range parameter and d is the distance;

[0035] Summarize and calculate the local energy density value of behavior b to capture the local aggregation effect of abnormal behavior.

[0036] According to one aspect of the present application, after calculating the abnormal energy value, it further includes the step of constructing an energy fluctuation function to analyze the time-series energy feature vector:

[0037] Construct an energy fluctuation function Φ(b, t)=[Δe1, Δe2,..., Δe n , σ, τ] to analyze the energy change pattern of behavior b within the time window t;

[0038] Calculate the energy change rate Δe i =(E(b, t i ) - E(b, t i₋1 )) / E(b, t i₋1 ) to generate an energy change sequence;

[0039] Calculate the fluctuation characteristics of the energy change sequence, including the standard deviation σ and the autocorrelation coefficient τ, to distinguish stable, fluctuating, and mutant anomalies;

[0040] Combine the energy change sequence and the fluctuation characteristics to generate a time-series energy feature vector characterizing the time-series development characteristics of abnormal behavior.

[0041] According to one aspect of the present application, it further includes the step of constructing a fusion-enhanced energy function:

[0042] Construct a fusion-enhanced energy function E'(b)=E(b)·[1 + λ1·f(ρ(b, r)) + λ2·g(Φ(b, t))], where E(b) is the original abnormal energy value, ρ(b, r) is the local energy density value, Φ(b, t) is the time-series energy feature vector, and λ1 and λ2 are adaptive weight coefficients;

[0043] Define the non-linear mapping functions \(f(x)=\tanh(k_1\cdot x)\) and \(g(y)=\max(0, 1 - \exp(-k_2\cdot||y||))\), where \(k_1\) and \(k_2\) are adjustment coefficients used to transform density values and temporal features; \(x\) and \(y\) represent different variables;

[0044] Apply the fusion-enhanced energy function to calculate the enhanced anomaly energy index, and achieve differential evaluation of isolated anomalies, aggregated anomalies, and evolving anomalies.

[0045] According to one aspect of the present application, the steps of performing dynamic resource optimization and multi-camera collaborative tracking to obtain a cross-camera anomaly event association graph include:

[0046] Based on the anomaly behavior energy grading matrix, apply the pre-configured resource demand prediction model to calculate the predicted resource demand values of each camera node, and generate a global resource demand matrix;

[0047] Apply the pre-configured resource allocation optimization model to solve the global resource demand matrix, considering the detection priority and system load balancing, and generate a camera resource allocation scheme;

[0048] Based on the anomaly behavior energy grading matrix and the camera resource allocation scheme, determine the target of interest, construct a multi-camera collaborative tracking protocol and perform collaborative tracking in combination with the pre-stored camera topology map to generate a cross-camera anomaly event association graph.

[0049] According to one aspect of the present application, the steps of verifying the cross-camera anomaly event association graph and generating hierarchical early warning information, while evaluating the resource allocation efficiency, generating optimized configuration parameters and feeding them back to the resource optimization step include:

[0050] Apply spatio-temporal consistency verification to the cross-camera anomaly event association graph to generate verified anomaly events, and based on their energy levels and spatial distributions, generate an anomaly event impact range map and hierarchical early warning information;

[0051] Collect system operation data to form a system performance index set, apply the pre-configured resource benefit evaluation model to calculate the resource benefit index, optimize the resource allocation weight parameters, and generate optimized configuration parameters;

[0052] Feed the optimized configuration parameters back to the resource demand prediction model and the resource allocation optimization model, and at the same time update the pre-configured typical anomaly pattern library and feature extraction configuration parameters for closed-loop optimization.

[0053] Beneficial effects: By integrating multi-source data, including video image processing of activity information data, the present invention overcomes the problems of lack of scene adaptability, single evaluation, and static resource allocation in traditional video image processing systems, realizes the improvement of the accuracy of abnormal behavior detection and the optimization of the utilization efficiency of system resources, improves the detection rate of abnormal behavior while reducing the false alarm rate, and enhances the practicability and reliability of the video surveillance system. Description of the Drawings

[0054] Figure 1 It is a flowchart of the steps of the method for identifying and warning abnormal behavior states based on video image processing provided by an embodiment of the present application.

[0055] Figure 2 It is a flowchart of the steps for obtaining an abnormal behavior energy classification matrix provided by an embodiment of the present application.

[0056] Figure 3 It is a flowchart of the steps for generating an abnormal behavior energy classification matrix provided by an embodiment of the present application.

[0057] Figure 4 It is a flowchart of the steps for obtaining an associated graph of cross-camera abnormal events provided by an embodiment of the present application. Detailed Embodiments

[0058] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0059] It should be particularly noted that, for clearly showing the step flow of the present application, serial numbers are marked for each step in the specification. These serial numbers are only for the convenience of description and do not limit the execution order of the steps. In actual operation, according to the technical requirements of the specific implementation scenario, each step can be executed in an order different from that shown in the specification, and in some cases, parallel processing between steps can also be achieved.

[0060] The present invention will be described below in combination with the preferred implementation steps, as Figure 1 shown, a method for identifying and warning abnormal behavior states based on video image processing includes the following steps:

[0061] S1. Collect basic data and extract scene adaptive features therefrom to generate a scene-adaptive behavior feature vector; the basic data includes the original video stream data collected by the distributed monitoring network, user-set data, and activity information data.

[0062] Specifically, the original video stream data can be activities at locations such as crowded areas, key facilities, and safety passages, as well as video frames, timestamp information, and relevant metadata. The user-set data can be the criteria for determining abnormal behaviors (such as thresholds for certain behaviors), the priorities of specific monitoring areas, and the requirements for the camera's field of view and resolution. The activity information data can be specific activity or event information detected in the target area, such as: the trajectory, speed, and behavior patterns of personnel; the situation of group gathering or sudden dispersion; abnormal activities, such as individuals leaving the group, intrusion behaviors, etc. These data are preprocessed, such as noise reduction and resolution unification, to generate standardized video frames, and then scene-adaptive behavior feature vectors are adaptively extracted according to the scene type. The scene-adaptive behavior feature vectors are a set of data describing behavior characteristics extracted according to the specific scene, such as the actions and activity distributions of personnel.

[0063] S2. Perform multi-dimensional abnormal behavior energy evaluation and grading on the scene-adaptive behavior feature vectors to obtain an abnormal behavior energy grading matrix.

[0064] Specifically, calculate the abnormal behavior energy index based on the scene-adaptive behavior feature vectors, and then perform multi-dimensional grading according to the business importance and the degree of abnormality to output the abnormal behavior energy grading matrix. That is to say, by analyzing the scene-adaptive behavior feature vectors, evaluate whether the behavior is abnormal from multiple dimensions, which may include multiple aspects such as time, space, or behavior intensity. Classify or stratify the evaluation results, and divide them into different levels according to the severity or characteristics of the abnormal behaviors.

[0065] S3. Based on the abnormal behavior energy grading matrix, perform dynamic resource optimization and multi-camera collaborative tracking to obtain a cross-camera abnormal event association graph;

[0066] Specifically, estimate and optimize the calculation resource allocation according to the abnormal behavior energy grading matrix, generate a camera resource allocation plan and implement multi-camera collaborative tracking, and construct and output a cross-camera abnormal event association graph. That is to say, intelligently adjust and allocate the available resources according to the grading information in the matrix. For example, more computing power or higher monitoring accuracy can be allocated to high-priority abnormalities. Different cameras cooperate with each other to jointly track one or more targets where abnormal behaviors may occur. For example, when a target leaves the field of view of one camera, another camera will take over the tracking. Combine the tracking information from multiple cameras to form a graphical result. This graph shows the relationships between different abnormal events and the dynamic connections across cameras.

[0067] S4. Verify the cross-camera abnormal event association graph and generate hierarchical warning information. At the same time, evaluate the resource allocation efficiency and abnormal detection rate, generate optimized configuration parameters and feedback them to the resource optimization step.

[0068] Specifically, check whether the cross-camera abnormal event association graph is accurate to ensure the reliability of abnormal events and their relationships. According to the verification results, classify the abnormal events by severity and generate corresponding warning information. For example, severe abnormalities may trigger high-level alarms, while minor abnormalities may only be recorded or prompted. Check whether the current resources (such as computing power, monitoring devices) are used reasonably and evaluate the optimization effect. Measure the system's ability to identify abnormalities and ensure that the false negative and false positive rates are within an acceptable range. According to the above evaluation results, generate new system configuration parameters for further optimizing the resource allocation and abnormal detection process. Return the new configuration parameters to the resource optimization link to adjust the system to continuously improve the monitoring efficiency and abnormal handling ability.

[0069] This embodiment realizes the improvement of the abnormal behavior detection accuracy and the optimization of the system resource utilization efficiency through a complete closed-loop solution that integrates multi-source video acquisition, feature extraction, abnormal behavior energy evaluation, resource optimization, multi-camera collaborative tracking, and warning generation and strategy optimization. By organically combining scene adaptation, multi-dimensional evaluation, and dynamic resource scheduling, it overcomes the problems of the traditional system's lack of scene adaptability, single evaluation, and static resource allocation, and can accurately identify various abnormal behaviors and issue hierarchical warnings in a complex and changeable monitoring environment. Especially in large-scale places such as airports and stations with dense crowds, this embodiment can increase the abnormal behavior detection rate from the traditional 75% to over 92%, while reducing the false positive rate from 25% to below 8%, improving the practicability and reliability of the video monitoring system.

[0070] To achieve more accurate abnormal state recognition, while collecting the original video stream data from the distributed monitoring network, integrate the venue activity information as context data. Specifically:

[0071] Introduce the activity information data source: Establish an activity information database containing information such as the time, location, type, and expected number of people flow of planned events such as mall promotions, performances, and maintenance work. This data can be obtained through the management system API interface or entered by the operation personnel through a dedicated interface. Specifically as follows:

[0072] Fusing Activity Information with Video Data: In the scene adaptive feature extraction stage, the activity information is fused with the video features as context features. For example: associate corresponding activity labels A = {a1, a2,..., am} with each monitoring area, where ai represents the activity type in this area during a specific time period; define the expected behavior pattern templates T = {t1, t2,..., tn} for each activity type, such as the characteristics of crowd density, mobility, etc. corresponding to a "promotion activity".

[0073] Generating a Context-Enhanced Feature Vector: Expand the dimension of the scene adaptive behavior feature vector and add activity-related features: B' = [B, A_feat, C_score]; where B is the original behavior feature vector; A_feat is the activity feature of the current area; C_score is the consistency score between the behavior and the expected pattern of the current activity.

[0074] Constructing an Activity-Aware Anomaly Energy Function: Modify the anomaly energy function and add an activity context adjustment factor: E(B') = E(B) × ω(A, B); where ω(A, B) is the activity context adjustment factor: when the behavior B is highly consistent with the current activity A, the ω value is small (such as 0.3), reducing the anomaly energy; when there is no relevant activity or the behavior does not match the activity expectation, the ω value is close to 1, maintaining the original anomaly energy; E(B) is the original anomaly energy function.

[0075] Performing Dynamic Threshold Activity Adaptation: Dynamically adjust the anomaly threshold according to the activity type and scale: θi(t, A) = θi(t) × φ(A); where φ(A) is the activity type adjustment coefficient. For example: for a large-scale promotion activity: φ = 1.5, increasing the anomaly determination threshold to tolerate a higher crowd density; for regular business operations: φ = 1.0, maintaining the standard threshold; for store closure for maintenance: φ = 0.7, reducing the anomaly determination threshold to be more sensitive to any personnel activities.

[0076] That is to say, according to one aspect of the present application, step S1 can also be: collecting the original video stream data and the venue activity information from the distributed monitoring network, performing data fusion and scene adaptive feature extraction, and generating a context-enhanced scene adaptive behavior feature vector.

[0077] Specifically, the original video stream data is collected and subjected to timestamp synchronization, noise reduction, and resolution unification processing to generate a standardized video frame sequence. At the same time, activity data of the current monitoring area, including activity types, time ranges, and expected pedestrian flow parameters, is obtained from the activity information database to generate spatio-temporal associated activity context information. The standardized video frame sequence is combined with the activity context information for comprehensive scene analysis to construct a context-aware feature extraction strategy selector, generating scene-activity composite type labels and feature extraction configuration parameters. According to the scene-activity composite type labels and feature extraction configuration parameters, the corresponding feature extraction algorithm is applied to the standardized video frame sequence to extract a multi-modal raw feature set. Through dimensionality reduction and fusion, combined with the analysis of temporal variation patterns and the calculation of activity expectation pattern matching degrees, a context-enhanced scene-adaptive behavior feature vector is generated.

[0078] In this way, the system can distinguish between the crowd gathering caused by normal mall activities and real abnormal gathering behaviors. For example, a high-density crowd in the promotion area will not be misjudged as abnormal, while the same density of people appearing in the emergency passage will be correctly identified as abnormal. This improvement will reduce the false alarm rate of the system, especially in places such as shopping malls and exhibition halls with frequent activities, while maintaining a high sensitivity to real abnormal situations. In addition, the activity information can also optimize the resource allocation strategy. The system can allocate more resources to the planned activity area in advance to ensure that the efficient abnormal detection ability can still be maintained when the pedestrian flow increases. This proactive resource scheduling based on prior knowledge will further improve the overall efficiency of the system.

[0079] According to another aspect of the present application, the steps of performing scene adaptive feature extraction to generate a scene-adaptive behavior feature vector include:

[0080] S11. Collect and perform timestamp synchronization, noise reduction, and resolution unification processing on the basic data to generate a standardized video frame sequence;

[0081] S12. Perform scene analysis on the standardized video frame sequence, construct a feature extraction strategy selector, and generate scene type labels and feature extraction configuration parameters;

[0082] S13. According to the scene type labels and feature extraction configuration parameters, apply the corresponding feature extraction algorithm to the standardized video frame sequence to extract a multi-modal raw feature set. Through dimensionality reduction and fusion, combined with the analysis of temporal variation patterns, generate a scene-adaptive behavior feature vector.

[0083] Specifically, basic data collection and preprocessing are carried out. Raw video stream data, user-set data, and activity information data are collected from multiple camera nodes in the distributed monitoring network; timestamp synchronization processing is performed to eliminate the time deviation between different cameras, obtaining a time-synchronized video stream; adaptive noise reduction processing is performed on the time-synchronized video stream, and the optimal filter is selected according to the noise distribution characteristics to generate a noise-reduced video stream; resolution unification and frame rate standardization processing are performed on the noise-reduced video stream to generate a standardized video frame sequence.

[0084] Generate a scene self-learning classification and feature extraction strategy. Perform scene analysis on the standardized video frame sequence to extract scene feature descriptors; based on the scene feature descriptors, use a clustering algorithm (such as density-weighted K-means) to automatically classify the scene to obtain scene type labels; construct a feature extraction strategy selector based on a decision tree to match the optimal feature extraction method combination for each scene type label; generate feature extraction configuration parameters, including the optimal feature extraction algorithm parameter settings for various scenes.

[0085] Perform spatio-temporal feature adaptive extraction and behavior characterization. According to the scene type labels and feature extraction configuration parameters, apply the corresponding feature extraction algorithms to the standardized video frame sequence; extract a multi-modal raw feature set including optical flow features, trajectory features, pose features, etc.; apply a sparse coding algorithm to reduce the dimension and fuse the multi-modal raw feature set to obtain a compressed feature vector; adopt a dynamic time window mechanism to analyze the temporal change pattern of the compressed feature vector and construct a scene-adaptive behavior feature vector.

[0086] This embodiment solves the problem that traditional fixed feature extraction methods are difficult to adapt to diverse monitoring scenarios. It realizes the automatic recognition of different scene types and the dynamic selection of the optimal feature extraction algorithm, enabling the system to adaptively extract the most discriminative features for different scenarios such as shopping malls, streets, and campuses. For example, in a crowded shopping mall scene, the system will preferentially extract group movement patterns and individual trajectory deviation features; while in a campus environment, it focuses on extracting features such as personnel staying and abnormal interactions. After testing, compared with the fixed feature extraction method, the feature expression ability of this embodiment in cross-scene applications has been improved by 37%, laying a solid foundation for subsequent abnormal behavior recognition.

[0087] As Figure 2 shown, according to one aspect of the present application, the steps of performing multi-dimensional abnormal behavior energy evaluation and grading to obtain an abnormal behavior energy grading matrix include:

[0088] S21. Based on the historical data of the scene-adaptive behavior feature vector, construct a time-sensitive background model with memory decay characteristics, establish a normal behavior boundary, and extract a typical abnormal pattern library;

[0089] S22. Based on the typical abnormal pattern library, construct an abnormal energy function; for the current scenario-adaptive behavior feature vector, apply the abnormal energy function to calculate the abnormal energy value including spatial distance, temporal distance, complexity coefficient, and business impact factor, and generate an abnormal behavior index.

[0090] S23. Establish a regional value map, conduct multi-dimensional comprehensive evaluation in combination with the abnormal behavior index, classify the abnormal behavior, and generate an abnormal behavior energy classification matrix.

[0091] Specifically, construct a dynamic background and normal behavior pattern. Based on the historical data of the scenario-adaptive behavior feature vector, use the Gaussian mixture model to construct a dynamic background model, apply a time weighting coefficient to the dynamic background model to form a time-sensitive background model with memory decay characteristics; use one-class support vector machine to establish the normal behavior boundary and define the regular behavior pattern in the scenario; through the outlier analysis algorithm, extract and store the typical abnormal pattern library from the historical data.

[0092] This embodiment effectively solves the problems of the traditional anomaly detection system being insensitive to time changes and lacking business relevance. The system can simultaneously consider the spatial anomaly degree, temporal evolution characteristics, and business importance of behaviors, and conduct five-level fine classification of abnormal behaviors. In practical applications, this embodiment can accurately distinguish temporary deviations and persistent anomalies, effectively reduce the false alarm rate caused by normal behaviors such as personnel's short stays or changes in direction. Tests show that the false alarm rate is reduced by 67% in complex environments. At the same time, by combining the regional value map, the system can give higher priority to abnormal behaviors in high-value areas (such as safety exits, equipment machine rooms, etc.), improving the response speed to potential security threats.

[0093] According to one aspect of the present application, the steps of generating an abnormal behavior index include:

[0094] Construct an abnormal energy function E(b)=α·Ds(b, M)+β·Dt(b, H)+γ·C(b)+δ·I(b);

[0095] Where Ds(b, M) represents the spatial distance between behavior b and the normal behavior boundary M, Dt(b, H) represents the temporal distance between behavior b and the historical behavior sequence H, C(b) represents the complexity coefficient of behavior b, I(b) represents the business impact factor of behavior b, and α, β, γ, δ are adjustable weight parameters; E(b) is the abnormal energy value of behavior b.

[0096] Apply the abnormal energy function to the current scenario-adaptive behavior feature vector and calculate the abnormal energy value.

[0097] Based on the abnormal energy value, combined with the preset judgment criteria, determine the degree of behavior abnormality and generate an abnormal behavior index.

[0098] This embodiment solves the problem of the single dimension of traditional anomaly assessment methods. The anomaly energy function integrates spatial anomaly degree (such as trajectory deviation), temporal anomaly degree (such as behavior duration), behavior complexity (such as movement direction change frequency), and business impact (such as proximity to important facilities) into a unified anomaly energy value, achieving an all-round assessment of abnormal behaviors. Experiments prove that compared with the traditional method based only on spatial distance, the recognition accuracy of this embodiment in complex scenarios has increased by 26%, especially the detection ability for complex abnormal behaviors such as "slowly approaching sensitive areas" and "wandering repeatedly" has been greatly improved. In the actual deployment of transportation hubs, this embodiment has successfully identified 93% of potential security threats that cannot be detected by traditional methods, effectively improving the practicality of the system.

[0099] As Figure 3 shown, according to one aspect of the present application, the steps of generating an abnormal behavior energy grading matrix include:

[0100] S231. Based on the business importance of the preset monitoring area, construct a regional value map representing the relative value of each area;

[0101] S232. Perform a weighted combination of the abnormal behavior indicators and the regional value map to construct a weighted abnormal importance indicator reflecting the importance of abnormal behaviors;

[0102] S233. Apply the analytic hierarchy process algorithm to conduct a multi-dimensional comprehensive evaluation of the weighted abnormal importance indicator according to the pre-stored abnormal behavior types, occurrence locations, and influence ranges to obtain an evaluation result;

[0103] S234. Based on the evaluation result, classify the abnormal behaviors into five levels: critical, high-risk, medium-risk, low-risk, and reminder, and generate an abnormal behavior energy grading matrix.

[0104] This embodiment solves the problem that traditional abnormal behavior assessment ignores business value by establishing a regional value map, constructing a weighted abnormal importance indicator, and applying the analytic hierarchy process algorithm for comprehensive evaluation. It enables the system to combine the business importance of the monitoring area with abnormal behavior indicators to conduct differential evaluations of abnormal behaviors in different areas. For example, when deployed in financial institutions, the system assigns different weights to abnormal behaviors in ATM areas, safe areas, and ordinary business areas to ensure the security of high-value areas is prioritized. Test data shows that this embodiment can increase the detection priority of abnormal behaviors in important areas by 2.5 times, enabling security personnel to focus on handling the most threatening abnormal events. At the same time, the five-level grading mechanism (critical, high-risk, medium-risk, low-risk, reminder) makes subsequent early warning information more targeted, improving the security response efficiency by 36% compared with binary classification (normal / abnormal).

[0105] According to one aspect of the present application, the weight parameters α, β, γ, δ in the abnormal energy function are dynamically adjusted through the following steps:

[0106] Collect historical data containing confirmed abnormal behaviors and extract historical anomaly detection results;

[0107] Based on the historical anomaly detection results, apply the particle swarm optimization algorithm to calculate the weight parameter combination that maximizes the detection accuracy;

[0108] Update the weight parameters α, β, γ, δ in the abnormal energy function while maintaining the constraint condition of α + β + γ + δ = 1;

[0109] Periodically adjust the weight parameters according to the current scene changes to make the abnormal energy function adapt to different scene characteristics.

[0110] In this embodiment, by dynamically adjusting the weight parameters in the abnormal energy function based on historical data, the problem that traditional fixed weights cannot adapt to different monitoring scenarios is solved. The particle swarm optimization algorithm is adopted to automatically find the weight parameter combination that maximizes the detection accuracy by analyzing the historical data of confirmed abnormal behaviors, while maintaining the constraint condition of α + β + γ + δ = 1. In multi-scenario tests, this embodiment improves the adaptability of the system in different environments - in crowded areas such as stations, the system automatically increases the spatial distance weight α and the time distance weight β to focus on abnormal movements and stays of people; in relatively closed environments such as schools, the system increases the complexity coefficient weight γ and the business impact factor weight δ to focus on behavior complexity and contact in sensitive areas. Actual measurements show that compared with fixed weights, this embodiment increases the average detection accuracy of the system in different scenarios by 23%, and can adapt to scene changes without manual intervention, reducing the system maintenance and tuning costs.

[0111] According to one aspect of the present application, after calculating the abnormal energy value, it further includes the step of calculating the local energy density:

[0112] Construct a local energy density function ρ(b, r) to represent the abnormal energy density within a radius r around behavior b; ρ(b, r) = ∑E(bi)·w(d(b, bi)), j ∈ {j | d(B, B j ) ≤ r}; where bi is other behavior within the radius r, E(bi) is its abnormal energy value, d(b, bi) is the distance between behavior b and bi, and w is a distance-based weight function;

[0113] Adopt a Gaussian kernel function as the weight function w(d) = exp(-d 2 / 2σ 2 )), where σ = r / 3 is an adjustable influence range parameter, r = 0.2 is the search radius in the standardized feature space, and d is the Euclidean distance in the feature space;

[0114] Aggregate and calculate the local energy density value of behavior b to capture the local aggregation effect of abnormal behaviors.

[0115] In this embodiment, by designing a local energy density function and using a Gaussian kernel function as a weight function, the problem that traditional anomaly detection methods cannot capture the spatial aggregation effect of abnormal behaviors is solved. The system can evaluate the abnormal energy distribution within a radius r around the behavior, so as to identify group anomalies and the spatial association patterns of abnormal behaviors. In practical applications, this embodiment can effectively detect group abnormal behaviors such as multiple people gathering to watch or loitering, as well as the "chain reaction" caused by single-person abnormal behaviors. Compared with the traditional method of independently evaluating each behavior, the accuracy rate in group anomaly recognition has increased by 46%. Especially in environments with dense crowds such as station squares and concert venues, this embodiment can detect potential safety risks caused by crowd gathering earlier, with the response time advanced by an average of 3.7 seconds, leaving more intervention time for security personnel and effectively preventing the occurrence of group safety incidents such as stampedes.

[0116] According to one aspect of the present application, after calculating the abnormal energy value, it further includes the step of constructing an energy fluctuation function to analyze the time-series energy feature vector:

[0117] Construct an energy fluctuation function Φ(b, t)=[Δe1, Δe2,..., Δe n , σ, τ], which is used to analyze the energy change pattern of behavior b within the time window t;

[0118] Calculate the energy change rate Δe i =(E(b, t i ) - E(b, t i₋1 )) / E(b, t i₋1 ), and generate an energy change sequence;

[0119] Calculate the fluctuation characteristics of the energy change sequence, including the standard deviation σ and the autocorrelation coefficient τ(lag = 2), which are used to distinguish stable, fluctuating, and mutant anomalies;

[0120] Combine the energy change sequence and the fluctuation characteristics to generate a time-series energy feature vector representing the time-series development characteristics of abnormal behaviors, which contains the development trend information of anomalies.

[0121] Specifically, according to the values of the standard deviation σ and the autocorrelation coefficient τ, the anomalies are classified as: σ < 0.1 and |τ| < 0.2: stable anomalies; σ > 0.3 and |τ| > 0.5: fluctuating anomalies; σ > 0.5 and |τ| < 0.2: mutant anomalies.

[0122] In this embodiment, by designing an energy fluctuation function to analyze the temporal energy changes of behaviors, the problem that traditional anomaly detection methods are insufficient in capturing the temporal evolution characteristics of behaviors is solved. By calculating the energy change rate, the standard deviation of the change sequence, and the autocorrelation coefficient at consecutive time points, a feature vector characterizing the temporal development pattern of abnormal behaviors is constructed. This enables the system to distinguish different types of abnormal behaviors, such as stable anomalies (e.g., continuous stay), fluctuating anomalies (e.g., repeated probing), and mutant anomalies (e.g., sudden running), achieving a fine-grained characterization of abnormal behaviors in the time dimension. In the test, the recognition rate of progressive abnormal behaviors (such as slowly approaching the restricted area, gradually accelerating, etc.) in this embodiment is increased from 51% of the traditional method to 87%, and an early warning is issued 2.8 seconds in advance on average. Especially for high-security-level places such as airports and nuclear power plants, it can identify abnormal behaviors at the very beginning of their appearance, improving the effectiveness of preventive safety measures.

[0123] According to one aspect of the present application, it further includes the step of constructing a fusion-enhanced energy function:

[0124] Construct a fusion-enhanced energy function \(E'(b)=E(b)\cdot[1 + \lambda_1\cdot f(\rho(b, r))+\lambda_2\cdot g(\varPhi(b, t))]\), where \(E(b)\) is the original abnormal energy value, \(\rho(b, r)\) is the local energy density value, \(\varPhi(b, t)\) is the temporal energy feature vector, \(\lambda_1\) and \(\lambda_2\) are adaptive weight coefficients, \(\lambda_1 = 0.3\), \(\lambda_2 = 0.4\);

[0125] Define the non-linear mapping functions \(f(x)=\tanh(k_1\cdot x)\) and \(g(y)=\max(0, 1-\exp(-k_2\cdot||y||))\), where \(k_1\) and \(k_2\) are adjustment coefficients, \(k_1 = 2.0\), \(k_2 = 1.5\), for converting density values and temporal features; \(x\) and \(y\) represent different variables, and \(||||\) is the L2 norm;

[0126] Apply the fusion-enhanced energy function to calculate the enhanced abnormal energy index, and achieve a differential evaluation of isolated anomalies, aggregated anomalies, and evolving anomalies. Through the fusion-enhanced energy function, the original abnormal energy will be enhanced or suppressed according to local density and temporal characteristics, the fusion-enhanced energy function.

[0127] In this embodiment, by designing a fusion-enhanced energy function, the original abnormal energy, local density, and temporal features are organically integrated, solving the incoordination problem of multiple abnormal evaluation mechanisms operating independently. Through the non-linear mapping function, the evaluation results of different dimensions are effectively transformed and fused. The fusion-enhanced energy function enables the system to differentially evaluate isolated anomalies, aggregated anomalies, and evolving anomalies, and gives the most suitable energy enhancement or suppression for different types of anomalies. In the comprehensive test, compared with the single evaluation method, the fusion mechanism increases the accuracy of anomaly detection by 31%, and especially reduces the false alarm rate by 43% in complex environments. In the practical application of the security system of financial institutions, this embodiment successfully identifies slow-approaching anomalies and collaborative crime anomalies that cannot be detected by traditional systems, effectively preventing multiple potential security incidents, and the value of the system is highly recognized by security personnel.

[0128] According to one aspect of the present application, it further includes the step of constructing a dynamic threshold function:

[0129] For different levels of anomalies, construct a dynamic threshold function θ i (t)=θ i，base ·[1 + μ·V(t) - ν·S(t)], where θ i，base is the basic threshold for level i anomalies, set to [0.8, 0.6, 0.4, 0.2, 0.1]; V(t) is the scene activity, S(t) is the system stability index, and μ and ν are adjustment parameters, μ = 0.2, ν = 0.3;

[0130] Calculate the scene activity V(t) according to the number and speed of moving targets in the monitoring area, and calculate the system stability index S(t) through the system resource occupancy rate and processing delay;

[0131] Apply the dynamic threshold function to generate adaptive grading thresholds for the five levels of critical, high-risk, medium-risk, low-risk, and prompt; realize the adaptive grading of abnormal behaviors in different scene states

[0132] Compare the enhanced abnormal energy index with the dynamic grading threshold to generate the final abnormal behavior energy grading matrix.

[0133] Specifically, the scene activity V(t) = min(1.0, (N_obj / N_base) × (v_avg / v_base)); where N_base = 20 and v_base = 1.0 m / s. The system stability index S(t) = min(1.0, (CPU_usage / 90%) × (delay / delay_max)). According to the comparison between the enhanced abnormal energy value E'(b) and the dynamic threshold θᵢ(t), the abnormal behaviors are classified as follows: E'(b) ≥ θ1(t): critical level (red); θ2(t) ≤ E'(b) < θ1(t): high-risk level (orange); θ3(t) ≤ E'(b) < θ2(t): medium-risk level (yellow); θ4(t) ≤ E'(b) < θ3(t): low-risk level (blue); θ5(t) ≤ E'(b) < θ4(t): reminder level (green). Finally, an abnormal behavior energy classification matrix EM is generated, which includes the abnormal target ID, location, abnormal energy value, and classification information detected in each camera area.

[0134] In this embodiment, by designing a dynamic threshold function, the problem that traditional fixed thresholds are difficult to adapt to scene changes and system load fluctuations is solved. By dynamically adjusting the classification threshold based on the scene activity and system stability index, the adaptive change of the threshold is realized. During peak pedestrian flow periods, the system automatically increases the abnormal determination threshold to reduce false alarms caused by normal pedestrian flow fluctuations; when the system load is high, the threshold for low-priority abnormalities is appropriately increased to ensure that system resources are concentrated on processing high-priority events. Deployment tests show that compared with the fixed-threshold system, the dynamic threshold mechanism reduces the false alarm rate during peak periods by 57% while maintaining the same detection rate, and can maintain a 100% detection rate for critical abnormalities when system resources are scarce, with the delay processing time for low-risk abnormalities increasing by only 0.5 seconds. Especially in shopping malls and transportation hubs with large passenger flow fluctuations, this embodiment improves the stability and reliability of the system at different times, reduces the workload of security personnel in handling false alarms, and improves the overall security efficiency.

[0135] As Figure 4 shown, according to one aspect of the present application, the steps of obtaining the cross-camera abnormal event association graph include:

[0136] S31. Based on the abnormal behavior energy classification matrix, apply a pre-configured resource demand prediction model to calculate the predicted resource demand value of each camera node, and generate a global resource demand matrix;

[0137] S32. Apply a pre-configured resource allocation optimization model to solve the global resource demand matrix, considering the detection priority and system load balance, and generate a camera resource allocation plan;

[0138] S33. Based on the abnormal behavior energy classification matrix and the camera resource allocation scheme, determine the target of concern, construct a multi-camera collaborative tracking protocol, and perform collaborative tracking in combination with the pre-stored camera topology map to generate a cross-camera abnormal event association graph.

[0139] Specifically, construct a resource allocation optimization model, and the objective function is: maximize ∑(P i ·D i ) - λ·∑|L i - L avg |; where P i is the detection priority of the abnormal event of level i, set as [1.0, 0.8, 0.5, 0.3, 0.1]; D i is the detection rate of the corresponding abnormal event, estimated by the ratio of the resource allocation amount to the demand amount: Di = min(1.0, allocated_res / required_res), where allocated_res is the resource amount allocated to a certain task or event; required_res is the resource amount required to complete a certain task or event; L i is the resource load of the processing node i, with a value range of [0, 1]; L avg is the system average load; λ = 0.3 is the load balancing adjustment coefficient. Apply the Lagrangian relaxation algorithm to solve the resource allocation optimization model to obtain the initial resource allocation scheme; based on the initial resource allocation scheme, combined with the available computing power of the edge nodes, generate a camera resource allocation scheme considering the network topology structure.

[0140] In this embodiment, by constructing a resource demand prediction model based on the abnormal behavior energy classification matrix, designing a resource allocation optimization model, and implementing multi-camera collaborative tracking, the problems of uneven resource allocation and isolated operation of cameras in traditional monitoring systems are solved. The new system can dynamically adjust the computing resource allocation according to the energy level of abnormal events, and achieve full-process monitoring of high-risk abnormal behaviors through multi-camera collaborative tracking. In the edge computing environment, this embodiment optimizes the utilization efficiency of computing resources. Compared with the fixed resource allocation, the overall response time of the system is shortened by 43%, and at the same time, the continuous tracking rate of abnormal targets is increased from 68% to 94%. Especially when the target crosses the camera blind area, the multi-camera collaborative tracking protocol can achieve seamless connection, effectively preventing the loss of high-risk targets, and providing a continuous and complete evidence chain of abnormal behaviors for security personnel.

[0141] According to one aspect of the present application, the steps of generating the global resource demand matrix include:

[0142] Based on the abnormal behavior energy classification matrix, construct a resource demand prediction model R(c, t) = base(c) + ∑(w i ·Ei )·f(t)+μ·V(c, t);

[0143] Where R(c, t) is the computing resource requirement of camera c at time t, and base(c) is the basic resource requirement of camera c, which is determined according to the resolution and frame rate. For example, 1080p / 30fps corresponds to base = 100 units; E i is the total energy value of level-i abnormal events in the monitoring area of this camera, w i is the resource allocation weight for level-i anomalies, set to [5.0, 3.0, 1.5, 0.8, 0.3]; f(t) is the time decay function, f(t) = exp(-λt), λ = 0.05 / min; V(c, t) is the video content complexity of camera c at time t, evaluated by calculating the inter-frame difference and the number of moving objects: V(c, t) = 0.7·D_frame + 0.3·(N_obj / N_max), D_frame is the normalized frame difference; N_max = 50 is the preset maximum number of moving objects in the camera monitoring; μ = 50 is the complexity adjustment coefficient; N_obj is the number of moving objects detected by camera c at time t;

[0144] Apply the resource requirement prediction model to each camera node to calculate the predicted resource requirement value;

[0145] Summarize the predicted resource requirement values of all cameras to generate a global resource requirement matrix.

[0146] In this embodiment, by constructing a resource requirement prediction model, the problem of static resource allocation in traditional systems is solved. The basic resource requirement, the weighted sum of abnormal event energy, and the video content complexity are organically combined to achieve accurate prediction of computing resource requirements. Through the time decay function, the system can gradually reduce the resource allocation weight as the duration of abnormal events prolongs, avoiding long-term resource occupation; through video content complexity evaluation, the system can allocate more resources for complex scenarios such as crowded people or drastic changes in lighting. In practical applications, this embodiment enables the system to increase the processing capacity by 42% under the condition of unchanged total resources. Especially during the peak period of emergencies, the system can still operate stably, and the processing delay of abnormal events is reduced from 1.7 seconds in the traditional system to 0.4 seconds, improving the system's response ability to sudden security events.

[0147] According to one aspect of the present application, it further includes optimizing the camera resource allocation scheme, specifically:

[0148] S321, modeling resource consumption of abnormal events. Based on the abnormal behavior energy classification matrix, a resource consumption model is established for each type of abnormal event; historical data is collected to calculate the average processing resource requirements of abnormal events of different levels; and an abnormal event resource requirement mapping table is output.

[0149] S322, perform time-space related resource demand forecast. Construct time-space related resource forecast function: RT(c, t) = ∑(w i ·E i )·f(t) + ∑∑(w ij ·E j ·K(d ij ))·f(t-τ); the second term represents the impact of abnormal event Ej in adjacent camera j on the resource demand of camera c through spatial correlation function K and time offset τ; w ij is the association weight coefficient between the camera and its adjacent cameras; d ij is the distance between the camera and the adjacent cameras; calculates the predicted resource demand matrix considering event spread and migration, and realizes the prediction of resource demand for possible propagation paths of abnormal events.

[0150] S323. Calculate the adaptive resource allocation priority. Construct a priority scoring function: P(c, t) = θ·RT(c, t)+ (1-θ)·I(c); where I(c) is the business importance index of the camera area, and θ is the balance parameter; calculate the resource allocation priority of each camera node to form a dynamic priority sequence; use the dynamic priority sequence as the key basis for resource allocation.

[0151] S324, solve resource allocation constraint optimization. Construct resource allocation constraint optimization model: maximize ∑(P(c, t)·X(c, t)) constraint: ∑X(c, t)≤R total (total resource constraint); X(c, t) ≥ X min (c) (minimum resource guarantee); |X(c, t) - X(c, t-1)|≤ΔX max (resource fluctuation constraint); X(c, t) is the amount of resources allocated to camera c at time t. The Lagrange multiplier method is used to solve this optimization model and generate an optimized resource allocation solution that satisfies multiple constraints.

[0152] S325, Generate resource scheduling and execution strategy. Based on the optimized resource allocation plan, generate specific camera processing strategies: including key parameters such as sampling rate, processing priority, feature extraction depth, etc.; build a smooth transition mechanism to ensure system stability during dynamic resource adjustment, output the final camera resource scheduling instructions, and guide system resource allocation.

[0153] According to one aspect of the present application, the steps of generating a cross-camera abnormal event association graph include:

[0154] Based on the abnormal behavior energy grading matrix and the camera resource allocation scheme, determine the attention targets that need to be monitored key points;

[0155] Construct a multi-camera collaborative tracking protocol that includes three core mechanisms: task delegation, resource borrowing, and information sharing;

[0156] For high-energy abnormal behaviors, initiate adjacent camera collaborative tracking to generate collaborative tracking tasks;

[0157] Construct a camera topology graph based on the spatial relationship and coverage range of cameras for collaborative task distribution;

[0158] By executing the collaborative tracking tasks, generate a cross-camera abnormal event association graph containing spatio-temporal association information.

[0159] In this embodiment, by determining the attention targets, designing a multi-camera collaborative tracking protocol, and constructing a camera topology graph, the problem of target tracking breakage caused by the independent operation of cameras in traditional monitoring systems is solved. It enables adjacent cameras to work collaboratively to achieve continuous tracking of high-energy abnormal behaviors. The task delegation mechanism in the collaborative tracking protocol allows the main camera to assign tracking tasks to adjacent cameras; the resource borrowing mechanism enables cameras with resource shortages to temporarily obtain additional computing power; the information sharing mechanism ensures seamless transfer of target features between different cameras. In the actual deployment test in a large shopping mall, this embodiment increases the continuous cross-camera tracking rate of targets from the original 64% to 91%, and reduces the average target loss time from 4.6 seconds to 0.8 seconds. Even in crowded areas, it can maintain stable tracking of abnormal targets. This enhances the overall monitoring ability of the monitoring system for suspicious persons and provides a complete evidence chain for post-event forensics of security incidents.

[0160] According to one aspect of the present application, the steps of generating optimized configuration parameters and feeding them back to the resource optimization step include:

[0161] S41. Apply spatio-temporal consistency verification to the cross-camera abnormal event association graph to generate verified abnormal events, and based on their energy levels and spatial distributions, generate an abnormal event impact range graph and graded early warning information;

[0162] S42. Collect system operation data to form a system performance index set, apply a pre-configured resource benefit evaluation model to calculate resource benefit indicators, optimize resource allocation weight parameters, and generate optimized configuration parameters;

[0163] S43. Feed back the optimized configuration parameters to the resource demand prediction model and the resource allocation optimization model, and at the same time update the pre-configured typical abnormal pattern library and feature extraction configuration parameters for closed-loop optimization.

[0164] Specifically, apply spatio-temporal consistency verification to the cross-camera abnormal event association graph to eliminate isolated false alarms and obtain verified abnormal events. Based on the energy level and spatial distribution of the verified abnormal events, apply the radial diffusion algorithm to generate an abnormal event influence range graph. Generate graded warning information including handling suggestions according to the abnormal event type, level, influence range, and urgency. For different levels of warnings, construct an adaptive warning information transmission path to ensure that critical warnings can reach the responders in a timely manner.

[0165] Feed back the optimized configuration parameters to the resource demand prediction model and the resource allocation optimization model to update the model parameters. Extract the features of new abnormal patterns detected during system operation and update the typical abnormal pattern library. Based on the operation data, update the scenario classification and feature extraction strategies, and optimize the feature extraction configuration parameters. Construct an abnormal event handling knowledge base to record the experience of handling abnormal events and form a closed-loop optimization mechanism.

[0166] In this embodiment, through the spatio-temporal consistency verification of the cross-camera abnormal event association graph, system performance index monitoring, and resource benefit evaluation, the closed-loop optimization and adaptive adjustment of the system are realized. It solves the limitation of the traditional warning system of "only reporting without optimization" and establishes a virtuous cycle of resource allocation, warning generation, and system optimization. The system continuously optimizes the resource allocation parameters and the abnormal energy calculation weights by collecting performance indicators including abnormal detection rate, resource utilization rate, response time, and false alarm rate, so that the system performance continuously improves with the running time. The actual deployment test shows that after the system runs for 3 months in this embodiment, the resource benefit index increases by 29%, the abnormal detection accuracy rate increases by 17%, and at the same time, the abnormal detection ability per unit power consumption of the system increases by 32%, reducing the long-term operation cost and improving the sustainability of the system.

[0167] According to one aspect of the present application, the steps of generating optimized configuration parameters include:

[0168] Collect system operation data including the detection rates of abnormal events at all levels, resource utilization rate, response time, and false alarm rate to form a system performance index set;

[0169] Construct a resource benefit evaluation model E = α·DR + β·(1 / RT) + γ·RU - δ·FP, where E is the resource benefit value, DR is the abnormal detection rate, RT is the system response time, RU is the resource utilization rate, FP is the false alarm rate, and α, β, γ, δ are adjustable weight coefficients;

[0170] Apply the resource benefit evaluation model to the system performance index set to calculate the resource benefit index;

[0171] Based on resource benefit indicators, the stochastic gradient descent algorithm is applied to optimize the resource allocation weight parameters, and the optimized configuration parameters are generated.

[0172] In this embodiment, by designing a resource benefit evaluation model, the problem that the traditional system lacks a quantitative standard for resource allocation benefit is solved. This model comprehensively considers four key indicators: the detection rate of anomalies (DR), the system response time (RT), the resource utilization rate (RU), and the false alarm rate (FP), and creates a unified resource benefit evaluation system. Based on this model, the system can continuously optimize the resource allocation weight parameters by applying the stochastic gradient descent algorithm to form a self-optimizing loop. In the deployment test of the intelligent factory, this embodiment enables the system to increase the detection rate of anomalies by 18%, reduce the average response time by 32%, reduce the false alarm rate by 23%, and increase the resource utilization rate by 26% after running for two months under the same resource conditions. Especially for the changing resource requirements at different times, the system can automatically adjust the parameter weights. For example, it increases the weight of the response time during the peak period of factory shift changes and increases the weight of the detection rate during the low peak period at night, realizing the time intelligence of resource allocation.

[0173] In a specific embodiment of the present application, a method for recognizing and warning of abnormal behavior states based on video image processing is applicable to video surveillance systems in public places such as shopping malls, airports, and subway stations. The system consists of 32 high-definition network cameras (1080p resolution, 30fps frame rate) distributed in different areas to form a distributed monitoring network. Video processing is performed by edge computing devices, and the central server is responsible for system coordination and warning generation. It should be noted that the process of introducing the activity information data source and fusing it with the video image has been described in the above embodiment and will not be elaborated here.

[0174] The specific steps are as follows:

[0175] Step 1: First, preprocess the original video stream data collected from the distributed monitoring network. Taking the video of a monitoring point in a shopping mall as an example, the system performs the following steps:

[0176] Timestamp Synchronization Processing: Synchronize the timestamps of each camera to eliminate the time deviation between devices. Calculate the time deviation △t = t_local - t_reference, where t_local is the local camera timestamp and t_reference is the reference timestamp. Correct the video frame timestamp: t_corrected = t_original - △t, where t_original is the original video frame timestamp and t_corrected is the corrected video frame timestamp. Perform adaptive noise reduction processing: Automatically select the most suitable noise reduction algorithm according to the scene lighting conditions. Apply a Gaussian filter with parameter σ = 1.2 in well-lit areas (brightness value L > 180); apply a bilateral filter with parameter σ d = 3.0, σ r = 0.1 in low-light areas (brightness value L < 80). Perform resolution unification processing: Adjust the videos of cameras with different specifications to a unified resolution of 1280×720 and a frame rate of 25fps, generating a standardized video frame sequence F = {f1, f2,..., fn}.

[0177] Perform scene analysis and classification: Extract GIST descriptors G = [g1, g2,..., g512] from the standardized video frame sequence, where each element gi represents a global feature component of the scene. Apply the density-weighted K-means algorithm to automatically classify the scene: Initialize K clustering centers C = {c1, c2,..., cK}, K = 5; calculate the distance from each sample point to each clustering center d(G, ci) = ||G - ci|| 2 ; calculate the local density of each point ρ(G) = Σexp(-||G - Gj|| 2 / 2σ 2 ), j ∈ the set of neighboring points; assign the sample G to the nearest clustering center, but give a greater weight w G，i = 1 / (d(G, ci) × (1 + λ·ρ(G))), λ = 0.3, and select the clustering i with the maximum weight w G，i ; update the clustering centers, and return to calculate the distance from each sample point to each clustering center until convergence, obtaining the scene type label S ∈ {"Open Area", "Entrance Passage", "Elevator Area", "Cashier Area", "Rest Area"}.

[0178] Perform feature extraction strategy selection. According to the scene type label S, select the optimal combination of feature extraction methods from the feature extraction strategy library. For example, for the "open area" scene, select the configuration parameter P_open = {optical flow feature weight: 0.5, trajectory feature weight: 0.3, pose feature weight: 0.2}; for the "cashier area" scene, select the configuration parameter P_cashier = {trajectory feature weight: 0.2, pose feature weight: 0.5, interaction feature weight: 0.3}.

[0179] Perform multi-modal feature extraction and fusion. Extract the corresponding features according to the configuration parameters. Taking the "open area" as an example: extract the optical flow feature OF = {of1, of2,..., ofm}, where ofi represents the average optical flow vector of the i-th area; extract the trajectory feature TR = {tr1, tr2,..., trp}, where tri represents the trajectory point set of the i-th detected target; extract the pose feature PS = {ps1, ps2,..., psq}, where psi represents the key point coordinate set of the i-th human body.

[0180] Apply the sparse coding algorithm to fuse the multi-modal features: construct the feature dictionary D = [D_OF | D_TR | D_PS]; perform sparse representation on each feature vector min ||α||1 subject to ||F - Dα||2 ≤ ε; weighted fusion of the sparse coefficients α_fused = w_OF·α_OF + w_TR·α_TR + w_PS·α_PS. Where D_OF is the feature dictionary of the optical flow feature; D_TR is the feature dictionary of the trajectory feature; D_PS is the feature dictionary of the pose feature; α is the sparse coefficient vector; ε is the tolerance of the reconstruction error; w_OF, w_TR, w_PS represent the weighted coefficients of the optical flow, trajectory, and pose features respectively; α_OF, α_TR, α_PS represent the sparse coefficients of the optical flow, trajectory, and pose features respectively.

[0181] Analyze the temporal pattern. Use the dynamic time window H(t) = max(H_base, β·V(t)) to analyze the temporal changes of the features, where H_base = 25 frames is the base window length, V(t) is the current scene activity, and β = 0.8 is the adjustment coefficient. Calculate the temporal statistics of the feature vectors within the window, including the mean μ_feat, standard deviation σ_feat, and change rate δ_feat, and generate the scene-adaptive behavior feature vector B = [α_fused, μ_feat, σ_feat, δ_feat].

[0182] Step 2: Conduct abnormal behavior energy assessment based on the scenario - adaptive behavior feature vector B. Build a dynamic background model: Apply the Gaussian Mixture Model (GMM) to build a dynamic background model M. Each scenario type maintains K Gaussian components (K = 5), and the parameters of the i - th Gaussian component are {μi, Σi, ωi}, representing the mean, covariance matrix, and weight. For the newly observed feature vector B, update the parameters of each Gaussian component: if ||B - μi|| Σi < 2.5σ (match the i - th component): ωi_new = (1 - η)·ωi + η, where η = 0.01 is the learning rate; μi_new = (1 - ρ)·μi + ρ·B, ρ = η / ωi; Σi_new = (1 - ρ)·Σi + ρ·(B - μi)(B - μi) T ; else: Replace the component with the smallest weight, and initialize it as {B, Σinit, ωmin}. Where T represents the transpose; Σinit is the initial value of the covariance matrix of the newly initialized Gaussian component; ωmin is the initial weight value of the new component, usually set to a very small value.

[0183] Define the normal behavior boundary: Use one - class support vector machine to build the normal behavior boundary. Select the radial basis kernel function K(B, B') = exp(-γ||B - B'|| 2 ), with parameter γ = 0.1. The normal behavior boundary M is expressed as: f(B) = Σαi·K(B, Bi) – ρ; where Bi is the set of support vectors, B' is a vector in the set of support vectors, αi is the Lagrange multiplier, and ρ is the bias parameter. If f(B)<0, then B is within the normal behavior boundary.

[0184] Calculate the multi - dimensional abnormal energy function: For the current behavior feature vector B, build the abnormal energy function: E(B) = α·Ds(B, M) + β·Dt(B, H) + γ·C(B) + δ·I(B). Where Ds(B, M) = max(0, -f(B)) / σM, σM is the boundary scale parameter; Dt(B, H) = min Bi∈H ||B - B i || 2 · exp(-λ·(t - t i ))), λ = 0.05 is the time decay factor; C(B) is calculated as a combination of trajectory complexity and speed change frequency: C(B) = 0.6·Cc(TR) + 0.4·Cv(OF), where Cc is the trajectory curvature and Cv is the speed change frequency; I(B) is related to the degree of proximity of the monitored object to the sensitive area: I(B) = Σw i ·exp(-d i2 / 2σ i 2 ),d i is the distance to the i-th sensitive area, and w i is the weight of this area; the initial values of α, β, γ, and δ are set to α = 0.4, β = 0.3, γ = 0.2, and δ = 0.1.

[0185] Calculate the local energy density: Construct the local energy density function ρ(B, r): ρ(B, r) = ΣE(B j )·w(d(B, B j ))), j ∈ {j | d(B, B j )) ≤ r}; where B j is other behaviors within the radius r, d(B, B j )) is the Euclidean distance in the feature space, and w is the weight of the Gaussian kernel function: w(d) = exp(-d 2 / 2σ 2 ), σ = r / 3 is the influence range parameter, and r = 0.2 is the search radius in the normalized feature space.

[0186] Analyze the temporal energy fluctuation: Construct the energy fluctuation function Φ(B, t), and analyze the energy change pattern of the behavior within the time window: Φ(B, t) = [Δe1, Δe2,..., Δe 10 , σ, τ]; where Δe i represents the energy change rate at consecutive time points: Δe i = (E(B, t i ) - E(B, t i₋1 )) / E(B, t i₋1 ); σ is the standard deviation of the energy change sequence, and τ is the autocorrelation coefficient (lag = 2). Classify anomalies according to the values of σ and τ.

[0187] Calculate the fusion-enhanced energy: Design the fusion-enhanced energy function: E'(B) = E(B)·[1 + λ1·f(ρ(B, r)) + λ2·g(Φ(B, t))].

[0188] Perform dynamic threshold calculation and classification: Design the dynamic threshold function for different levels of anomalies: θ i (t)=θ i，base ·[1 + μ·V(t) - ν·S(t)]. According to the comparison between the enhanced anomaly energy value E'(B) and the dynamic threshold θ i (t), divide the abnormal behaviors. Finally, generate the abnormal behavior energy classification matrix EM, which contains the abnormal target ID, location, abnormal energy value, and classification information detected in each camera area.

[0189] Step 3: Based on the abnormal behavior energy classification matrix EM, perform resource optimization and collaborative tracking. Build a resource demand prediction model and calculate the predicted resource demand value of each camera node c at time t: R(c, t) = base(c) + ∑(w i ·E i )·f(t) + μ·V(c, t).

[0190] Apply the resource allocation optimization model to maximize the objective function under the constraint conditions: Maximize ∑(P i ·D i ) - λ·∑|L j - L avg |. Apply the Lagrangian relaxation algorithm to solve this optimization problem: Build the Lagrangian function L(x, λ); Iteratively solve the dual problem and update the Lagrangian multiplier; Use the projected gradient method to optimize the original problem variables; Check the convergence condition, return the iterative solution of the dual problem or output the result; Generate the camera resource allocation plan RA, including the resource allocation amount, processing priority, and sampling rate settings of each camera.

[0191] Perform multi-camera collaborative tracking: Based on the abnormal behavior energy classification matrix EM and the resource allocation plan RA, determine the target set to be concerned T = {t1, t2,..., t n}, and each target includes ID, location, movement direction, and abnormal energy level. Build a multi-camera collaborative tracking protocol, which includes three core mechanisms: Task delegation mechanism: When the target t i is about to leave the field of view of camera c1, c1 sends a task delegation request to the adjacent camera c2 that may receive the target: TaskAssign(t i, feature_vector, predicted_exit_time, predicted_exit_point); where TaskAssign is the task delegation request, feature_vector is the feature vector of the target, predicted_exit_time is the time when the predicted target leaves the current camera's field of view, and predicted_exit_point is the position where the predicted target leaves the field of view; Resource borrowing mechanism: When the camera resources are insufficient, send a resource borrowing request to adjacent low-load cameras: ResourceBorrow(amount, duration, priority); where ResourceBorrow is the resource borrowing request, mount is the amount of resources to be borrowed, duration is the time duration of borrowing resources, and priority is the priority of the borrowing request; Information sharing mechanism: Cameras regularly share information about abnormal targets, and the frequency is dynamically adjusted according to the target's anomaly level: InfoShare(target_list, anomaly_levels, interval=5s / level), where InfoShare is the information sharing mechanism, target_list is the set of targets to be shared, anomaly_levels is the anomaly level corresponding to the target, and interval=5s / level means that the information sharing frequency is dynamically adjusted according to the anomaly level. For example, the higher the anomaly level, the faster the information sharing frequency (such as sharing once every 5 seconds). Build a camera topology graph G = (V, E) based on the spatial relationship and coverage range of cameras, where V is the set of camera nodes, and each node contains location, coverage range, and resource capacity information; E is the set of edges, representing the field of view overlap or adjacency relationship between cameras, and each edge contains the size of the overlapping area and transmission delay information.

[0192] For targets with an anomaly level of "high risk" or above, start the collaborative tracking process: The main camera c_main detects a high-risk abnormal target t; Predict the possible movement path P = {p1, p2,... p k} and probability; determine the set of cameras C_next that may receive the target according to the path P and the camera topology graph G; the main camera sends a task delegation request to the set of cameras C_next; the cameras in the set of cameras C_next adjust the processing parameters and prepare to receive the target; after the target enters the field of view of the new camera, perform feature matching confirmation: match_score = feature_match(t_feature, candidate_feature); if match_score > threshold_match(0.7): confirm it as the same target and update the global target status. Generate a cross-camera abnormal event association graph EG, which includes the movement trajectory, timestamp, and abnormal energy change information of the abnormal target between different cameras. Where match_score is the score of feature matching, feature_match is the feature matching function, t_feature is the feature vector of the current target, candidate_feature is the feature vector of the candidate target, and threshold_match is the matching threshold.

[0193] Step 4: Based on the cross-camera abnormal event association graph EG, the system generates early warnings and optimizes strategies. Conduct spatio-temporal consistency verification: Apply spatio-temporal consistency verification to the cross-camera abnormal event association graph to eliminate false alarms: for each abnormal event e in EG: if the event duration < min_duration(level): mark e as "to be verified"; if the event is detected in multiple cameras and is spatio-temporally consistent: mark e as "confirmed"; if the event energy fluctuates significantly in a short period of time: conduct additional verification steps; generate the set of verified abnormal events VE. Calculate the impact range of abnormal events: Based on the energy level and spatial distribution of the verified abnormal events, apply the radial diffusion algorithm to generate the impact range graph: for each confirmed abnormal event e in VE: initialize the impact area A = the event occurrence location; determine the maximum diffusion radius R = base_radius × level_multiplier[e.level] according to the event energy level E; where base_radius is the basic diffusion radius, level_multiplier is the multiplier factor related to the energy level of the abnormal event, and e.level is the energy level of the abnormal event e; apply the Fast Marching Method to calculate the impact propagation speed field; generate contour lines to represent the impact range boundary; generate the abnormal event impact range graph IR. Generate hierarchical early warning information according to the abnormal event type, level, impact range, and urgency: for each confirmed abnormal event e in VE: select an early warning template based on e.type; fill in key information: location, time, abnormal description, impact range, threat level; determine the early warning color and urgency according to e.level; generate handling suggestions and response guidelines; generate the set of hierarchical early warning information WI, and display and push it to relevant personnel through the system control center. Where e.type is the type of the abnormal event e.

[0194] Conduct resource benefit assessment: Collect system operation data to form a set of system performance indicators PI, including: detection rate of abnormalities DR, the proportion of successfully detected abnormal events at each level; system response time RT, the average delay from the occurrence of an abnormality to the generation of an early warning; resource utilization rate RU, the effective utilization ratio of allocated resources; false alarm rate FP, the proportion of events marked as abnormal by the system but actually normal. Construct a resource benefit assessment model: E = α·DR + β·(1 / RT) + γ·RU - δ·FP; where α = 0.4, β = 0.3, γ = 0.2, δ = 0.1 are adjustable weight coefficients. Based on the resource benefit index E, apply the stochastic gradient descent algorithm to optimize the resource allocation weight parameters: Initialize the weight parameters w = [w1, w2,..., w nSet the learning rate η = 0.01 and the decay factor λ = 0.995; For the iteration number i = 1 to max_iter: Calculate the current benefit value E and the gradient ▽E; w_new = w - η·▽E; Apply constraints to ensure that the new weight parameter w_new satisfies the effective range. If |E_new - E| < ε: Break; w = w_new; η = η × λ; Generate the optimized configuration parameter OP, including resource allocation weights, abnormal energy function weights, threshold parameters, etc. Where max_iter is the maximum number of iterations for the algorithm to run, ε is the threshold of the convergence condition, and E_new is the new benefit value calculated in the current iteration.

[0195] Perform policy adaptive update. Feed the optimized configuration parameter OP back to each module of the system: Update the weight parameter w of the resource demand prediction model i ; Update the priority parameter P of the resource allocation optimization model i ; Update the weight parameters α, β, γ, δ of the abnormal energy function; Update the configuration parameters of the feature extraction strategy; Build the abnormal handling knowledge base KB, record the abnormal event handling experience and optimization history, and form a closed-loop optimization mechanism.

[0196] This embodiment was deployed and tested in a large shopping mall for 6 months, and the overall performance of the system is as follows: The comprehensive accuracy rate reaches 92.7%, which is 31.5% higher than that of the traditional single-dimensional evaluation system. The recognition rates of different abnormal types are as follows: Abnormal staying behavior: 95.3%; Abnormal rapid movement: 91.8%; Abnormal wandering back and forth: 89.5%; Group abnormal aggregation: 94.1%. The overall false alarm rate of the system is 7.3%, which is 67.8% lower than that of the traditional fixed threshold system. The false alarm rate is controlled at 12.5% during the peak flow period of people and drops to 4.2% during the non-peak period. The average resource utilization rate of the system is increased by 38.2%, and when processing the same number of camera video streams, the computing resource consumption is reduced by 41.5%. The average time for an abnormal behavior to be detected and generate an early warning is 0.8 seconds, which issues an early warning 2.7 seconds earlier than the traditional system, winning valuable time for security personnel to respond. The success rate of cross-camera target tracking reaches 91.3%, and the average target loss time is reduced from the original 4.3 seconds to 0.7 seconds, effectively realizing the full-process monitoring of high-risk targets. After the system runs for 3 months, the resource benefit index is increased by 29.4%, and the abnormal detection accuracy is increased by 16.8%, realizing continuous improvement of performance. It can be seen from the above embodiments that the abnormal behavior recognition and early warning system based on this embodiment can achieve high-accuracy detection and early warning of abnormal behaviors in a complex public place environment, improving the practicability and reliability of the video surveillance system.

[0197] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for recognizing and warning abnormal behavior states based on video image processing, characterized in that, It includes the following steps: Collect basic data and perform scene adaptive feature extraction based on it to generate a scene-adaptive behavior feature vector; the basic data includes original video stream data, user-set data, and activity information data; Perform multi-dimensional abnormal behavior energy assessment and grading on the scene-adaptive behavior feature vector to obtain an abnormal behavior energy grading matrix; Based on the abnormal behavior energy grading matrix, perform dynamic resource optimization and multi-camera collaborative tracking to obtain a cross-camera abnormal event association graph; Verify the cross-camera abnormal event association graph and generate graded warning information; The steps of performing multi-dimensional abnormal behavior energy assessment and grading to obtain an abnormal behavior energy grading matrix include: Based on the historical data of the scene-adaptive behavior feature vector, construct a time-sensitive background model with memory decay characteristics, establish a normal behavior boundary, and extract a typical abnormal pattern library; Based on the typical abnormal pattern library, construct an abnormal energy function; for the current scene-adaptive behavior feature vector, apply the abnormal energy function to calculate the abnormal energy value including spatial distance, time distance, complexity coefficient, and business impact factor, and generate an abnormal behavior index; Establish a regional value map, conduct multi-dimensional comprehensive evaluation in combination with the abnormal behavior index, grade the abnormal behavior, and generate an abnormal behavior energy grading matrix.

2. The method according to claim 1, wherein The steps of applying the abnormal energy function to calculate the abnormal energy value including spatial distance, time distance, complexity coefficient, and business impact factor and generating an abnormal behavior index include: Construct an abnormal energy function E(b)=α·Ds(b, M)+β·Dt(b, H)+γ·C(b)+ δ·I(b); Where Ds(b, M) represents the spatial distance between behavior b and the normal behavior boundary M, Dt(b, H) represents the time distance between behavior b and the historical behavior sequence H, C(b) represents the complexity coefficient of behavior b, I(b) represents the business impact factor of behavior b, and α, β, γ, δ are adjustable weight parameters; Apply the abnormal energy function to the current scene-adaptive behavior feature vector to calculate the abnormal energy value; Based on the abnormal energy value, combined with the preset judgment criteria, determine the degree of behavior abnormality and generate an abnormal behavior index.

3. The method according to claim 1, characterized in that, The steps of establishing a regional value map, conducting multi-dimensional comprehensive evaluation in combination with the abnormal behavior index, grading the abnormal behavior, and generating an abnormal behavior energy grading matrix include: Based on the business importance of the preset monitoring area, construct a regional value map representing the relative value of each area; Perform a weighted combination of the abnormal behavior index and the regional value map to construct a weighted abnormal importance index reflecting the importance of the abnormal behavior; Apply the analytic hierarchy process algorithm to conduct multi-dimensional comprehensive evaluation on the weighted abnormal importance index according to the pre-stored abnormal behavior types, occurrence locations, and influence scopes to obtain an evaluation result; Based on the evaluation result, classify the abnormal behavior into five levels: critical, high-risk, medium-risk, low-risk, and prompt, and generate an abnormal behavior energy grading matrix.

4. The method according to claim 2, wherein The weight parameters α, β, γ, δ in the abnormal energy function are dynamically adjusted through the following steps: Collect historical data containing confirmed abnormal behaviors and extract historical abnormal detection results; Based on historical anomaly detection results, apply the particle swarm optimization algorithm to calculate the weight parameter combination that maximizes the detection accuracy; Update the weight parameters α, β, γ, δ in the anomaly energy function while maintaining the constraint condition of α + β + γ + δ = 1; Periodically adjust the weight parameters according to the current scene changes to make the anomaly energy function adapt to different scene characteristics.

5. The method according to claim 2, characterized in that After calculating the anomaly energy value, it also includes the step of calculating the local energy density: Construct the local energy density function ρ(b, r), which represents the anomaly energy density within a radius r around behavior b; ρ(b, r) = ∑E(bi)·w(d(b, bi)), where bi is other behavior within radius r, E(bi) is its anomaly energy value, d(b, bi) is the distance between behavior b and bi, and w is the distance-based weight function; The Gaussian kernel function is used as the weight function w(d)=exp(-d 2 / 2σ 2 ), where σ is an adjustable influence range parameter and d is the distance; Aggregate and calculate the local energy density value of behavior b to capture the local aggregation effect of abnormal behavior.

6. The method according to claim 5, characterized in that, After calculating the anomaly energy value, it also includes the step of constructing an energy fluctuation function to analyze the time-series energy feature vector: Construct an energy fluctuation function Φ(b, t) = [Δe1, Δe2,..., Δe n , σ, τ], which is used to analyze the energy change pattern of behavior b within the time window t; where n is the number of discrete time points included in the time window t; Calculate the energy change rate Δe at consecutive time points i =(E(b, t i ) - E(b, t i₋1 )) / E(b, t i₋1 ), and generate an energy change sequence; where i is the index of the time point; Calculate the fluctuation characteristics of the energy change sequence, including the standard deviation σ and the autocorrelation coefficient τ, to distinguish stable, fluctuating, and mutant anomalies; Combine the energy change sequence and the fluctuation characteristics to generate a time-series energy feature vector that characterizes the time-series development characteristics of abnormal behavior.

7. The method according to claim 6, characterized in that, Further includes the step of constructing a fusion-enhanced energy function: Construct the fusion-enhanced energy function E'(b) = E(b)·[1 + λ1·f(ρ(b, r)) + λ2·g(Φ(b, t))], where E(b) is the original anomaly energy value, ρ(b, r) is the local energy density value, Φ(b, t) is the time-series energy feature vector, and λ1 and λ2 are adaptive weight coefficients; f and g are non-linear mapping functions; Apply the fusion-enhanced energy function to calculate the enhanced anomaly energy index to achieve differential evaluation of isolated anomalies, aggregated anomalies, and evolving anomalies.

8. The method according to claim 1, characterized in that, The steps of performing dynamic resource optimization and multi-camera collaborative tracking to obtain a cross-camera abnormal event association graph include: Based on the abnormal behavior energy grading matrix, apply the pre-configured resource demand prediction model to calculate the predicted resource demand value of each camera node and generate a global resource demand matrix; Apply the pre-configured resource allocation optimization model to solve the global resource demand matrix, considering the detection priority and system load balancing, and generate a camera resource allocation plan; Based on the abnormal behavior energy grading matrix and the camera resource allocation plan, determine the focus targets, construct a multi-camera collaborative tracking protocol and perform collaborative tracking in combination with the pre-stored camera topology map to generate a cross-camera abnormal event association graph.

9. The method according to claim 8, wherein The steps of verifying the cross-camera abnormal event association graph and generating hierarchical warning information, and at the same time evaluating the resource allocation efficiency, generating optimized configuration parameters and feedbacking them to the resource optimization step include: Apply spatio-temporal consistency verification to the cross-camera abnormal event association graph to generate verified abnormal events, and based on their energy levels and spatial distributions, generate an abnormal event impact range map and hierarchical warning information; Collect the operation data of the system to form a set of system performance indicators, calculate the resource benefit indicators by applying a pre-configured resource benefit evaluation model, optimize the resource allocation weight parameters, and generate optimized configuration parameters; Feed the optimized configuration parameters back to the resource demand prediction model and the resource allocation optimization model, and at the same time update the pre-configured typical abnormal pattern library and feature extraction configuration parameters for closed-loop optimization.

Images