Multi-target identification sensor array safety monitoring method

Abstract

The invention relates to a multi-target identification sensor array safety monitoring method, and aims to improve the safety perception and target identification capability in a complex environment, and the method forms an array by deploying at least two different types of sensors such as acoustic, vibration, electromagnetic, infrared and optical sensors. After collected environment data and target signals are preprocessed, high-representativeness fusion feature vectors are generated through multi-dimensional feature extraction and a dynamic fusion strategy. And in combination with multi-level thresholding methods such as a dynamic threshold and a fixed threshold, a potential target signal is accurately extracted and a preliminary event is generated. A hierarchical storage mechanism is adopted, a large table and a small table are created, high-confidence events are quickly retrieved through the small table, complete context information is reserved in the large table, and discrete events are aggregated into composite events through space-time window association analysis. And a multi-target classifier based on machine learning / deep learning performs type identification and attribute analysis on the composite event, and finally completes threat assessment and early warning in combination with a preset security policy library.

Description

Technical Field

[0001] This application relates to the field of security monitoring technology, and in particular to a security monitoring method for a sensor array for multi-target identification. Background Technology

[0002] With the increasing demand for public safety and environmental monitoring, traditional single sensors or simple sensor networks face numerous challenges in multi-target identification and security monitoring in complex environments. While existing technologies, such as sensor arrays, can provide richer data, they still have shortcomings in data processing, feature fusion, target identification, and threat assessment.

[0003] Meanwhile, existing methods often use fixed threshold settings when extracting potential target signals from massive amounts of data, making it difficult to adapt to dynamically changing environments and prone to false positives or false negatives. For multi-target scenarios, the lack of effective event correlation and classification mechanisms makes it difficult to accurately identify and track multiple targets. Furthermore, simple data stitching or averaging fusion fails to fully utilize the complementarity and redundancy of different types of sensor data, resulting in limited feature representation capabilities and difficulty in effectively distinguishing multiple targets in complex environments.

[0004] Chinese patent CNCN115086616A discloses a method and system for identifying dangerous behaviors based on multiple sensors. The method includes receiving a monitoring request containing range information input by a user; determining a monitoring area based on the monitoring request; performing content recognition on the monitoring area; determining the monitoring frequency of each monitoring node based on the content recognition result; sending the monitoring frequency to each monitoring node; acquiring monitoring information based on each monitoring node; determining the display brightness of the monitoring point based on the monitoring information; and generating a visual area model based on the monitoring point containing the display brightness.

[0005] While the aforementioned solution can determine the monitoring area based on user requests, acquire monitoring information within the monitoring area according to preset monitoring nodes, identify problem areas, and provide real-time visualization of these problem areas, greatly reducing the workload of staff and indirectly improving the level of supervision, thus facilitating widespread adoption, the system cannot automatically adjust sensor operating parameters or data acquisition density when sudden anomalies occur in the environment or target behavior patterns change. This results in limited timeliness and completeness of key information capture. Furthermore, the system's decision-making mechanism relies excessively on manually preset rule bases, lacks autonomous learning capabilities when facing new and unknown threats, and fails to integrate statistical patterns from historical data for dynamic threshold optimization. Consequently, the consistency of judgment criteria for the same behavior is poor under different lighting, weather, or electromagnetic environments, leading to high false alarm and false negative rates.

[0006] Therefore, we propose a multi-target identification sensor array security monitoring method. Summary of the Invention

[0007] The main objective of this application is to provide a sensor array security monitoring method that can efficiently process multi-source heterogeneous sensor data, achieve accurate feature fusion, improve the accuracy of multi-target recognition, and support intelligent threat assessment.

[0008] To achieve the above objectives, this application provides a multi-target identification sensor array security monitoring method, comprising the following steps:

[0009] S1. A sensor array consisting of at least two different types of sensors collects environmental data and target signal data within the monitoring area. The sensor types include acoustic sensors, vibration sensors, electromagnetic sensors, infrared sensors, and optical sensors. The sensor array is arranged according to a predetermined grid, ring, or star topology. The spacing between adjacent sensors is determined based on the area of ​​the monitoring area and the required monitoring accuracy.

[0010] S2. Perform data preprocessing on the collected environmental data and target signal data to eliminate noise and standardize the data format;

[0011] S3. Perform feature extraction and fusion on the preprocessed sensor array data to generate a fused feature vector;

[0012] S4. Apply a thresholding method to the fused feature vector, which is any one or a combination of at least two of the following: dynamic threshold, fixed threshold, binarization threshold, and water immersion threshold, to extract potential target signals and generate preliminary target events.

[0013] S5. Create and store a large table and a small table based on the preliminary target event, wherein the large table stores the complete context information of the preliminary target event, and the small table stores short-term high-confidence summary information of the preliminary target event;

[0014] S6. Using the high-confidence event summary information in the small table as the core, and based on the timestamp, spatial stamp, and event type included therein, define a spatiotemporal window in the large table, and retrieve all preliminary target events within the window;

[0015] S7. Perform correlation analysis on the preliminary target events retrieved within the spatiotemporal window, and merge several preliminary target events belonging to the same target object into a composite event based on the analysis results, and mark the preliminary target events that do not belong to the same target object as a set of events to be observed.

[0016] S8. Based on a preset multi-target classifier, identify and classify the target type of the composite event, and output the type label and attribute information of each identified target;

[0017] S9. Based on the identification results and the preset security policy library, a threat assessment is performed. If a real threat is determined, a predefined alarm mechanism is triggered.

[0018] Preferably, in step S2, the denoising in the data preprocessing adopts any one or more combinations of wavelet transform denoising, median filtering denoising, and Kalman filtering denoising; the data preprocessing also includes at least one operation of random cropping, flipping, rotating, and scaling.

[0019] Preferably, in step S3, the feature extraction and fusion includes:

[0020] For the preprocessed data of each sensor node, features are extracted from multiple dimensions, including time domain, frequency domain, time-frequency domain, and spatial distribution, to form the initial feature vector of that node;

[0021] According to the preset fusion strategy, the initial feature vectors from different sensor nodes are integrated to generate the final fused feature vector. The fused feature vector includes a combination of at least two vectors among time domain features, frequency domain features, time-frequency domain features, and spatial distribution features.

[0022] Preferably, the preset fusion strategy includes:

[0023] When the complementarity and consistency between features are higher than the preset first complementarity threshold, and the feature dimension is higher than the preset second dimension threshold, a feature splicing and fusion strategy is adopted.

[0024] When the redundancy between features exceeds the preset third redundancy threshold, or the feature dimension exceeds the preset fourth dimension threshold, a principal component analysis dimensionality reduction and fusion strategy is adopted.

[0025] If the conditions for either of the above two strategies are not triggered, a dynamic weighted fusion strategy will be adopted.

[0026] Preferably, in step S4, the thresholding method employs a combination of dynamic thresholding and binarized thresholding, specifically including:

[0027] Dynamically calculate adaptive thresholds based on historical statistical values ​​of fused feature vectors;

[0028] The continuous feature vector is discretized into a 0 / 1 state sequence by using a binarization threshold;

[0029] When the binarized state continues for more than the time window set by the dynamic threshold, it is determined to be a valid target event.

[0030] Preferably, the formula for calculating the adaptive threshold is:

[0031] ;

[0032] Where μ represents the mean of a reference feature component within the historical time window, σ represents the standard deviation of the same reference feature component, and α is an adjustable sensitivity parameter.

[0033] Preferably, in step S4, the thresholding method employs a combination of immersion threshold and fixed threshold processing, specifically including:

[0034] A fixed threshold is used as the basic trigger condition;

[0035] The water immersion threshold method was used to perform regional growth analysis on the triggered signal;

[0036] When the coverage rate of the growth area reaches the preset ratio, it is determined to be a valid target event.

[0037] Preferably, in step S5:

[0038] The information stored in the large table includes: associated event ID, original data index, environmental context data, fused feature vector, and event processing status marker;

[0039] The information stored in the small table includes: event type identifier, timestamp, spatial stamp, confidence score, and trigger sensor code.

[0040] Preferably, the multi-objective classifier is constructed based on a machine learning model or a deep learning model, including one of a convolutional neural network, a recurrent neural network, a support vector machine, or an ensemble learning model.

[0041] Preferably, in step S9, the alarm mechanism includes one or more of the following: audible and visual alarm, work order push, system log recording, and linkage with other security systems.

[0042] The beneficial effects of the technical solution of this invention are as follows:

[0043] By employing a heterogeneous sensor array (acoustic / vibration / electromagnetic / infrared / optical combination) and adaptive topology layout, full-dimensional coverage and dynamic adjustment of the monitoring area are achieved. The complementarity of data from different frequency bands and modalities significantly improves target resolution in complex scenes. Simultaneously, by combining multiple sensor types and employing advanced feature extraction and fusion strategies, the complementarity and redundancy of different sensor data can be fully utilized to generate more representative fused feature vectors, significantly improving the ability to perceive multiple targets in complex environments. Furthermore, the introduction of various thresholding methods, such as dynamic thresholding and binarized thresholding, makes the extraction of potential target signals more flexible and accurate, effectively reducing false alarms and false negatives, and improving the quality of initial target event generation.

[0044] A hierarchical storage mechanism using large and small tables enables refined management of preliminary target events. The small table is used for quick retrieval of high-confidence events, while the large table retains complete contextual information. Combined with correlation analysis within the spatiotemporal window, it can efficiently aggregate scattered preliminary events into composite events, laying a solid foundation for multi-target recognition.

[0045] Multi-target classifiers based on machine learning or deep learning can accurately identify and classify target types for complex events, outputting detailed target attribute information. Combined with a pre-defined security policy library for threat assessment, this achieves intelligent operation across the entire chain from perception and identification to early warning, significantly improving the response speed and decision-making capabilities of the security monitoring system.

[0046] Each stage of the system, such as data acquisition, preprocessing, feature extraction, fusion, thresholding, and target recognition, adopts a modular design approach. This allows for easy addition, deletion, or replacement of modules based on specific monitoring scenarios and needs in practical applications. Thus, if a particular type of target in a specific area requires more precise monitoring, only targeted optimization or addition of relevant feature extraction modules is needed, without requiring large-scale modifications to the entire system. Attached Figure Description

[0047] Figure 1 This is a logical schematic diagram of a sensor array security monitoring method for multi-target identification in one embodiment of this application;

[0048] Figure 2 This is a flowchart of the feature extraction and fusion steps of a sensor array security monitoring method for multi-target identification in one embodiment of this application;

[0049] Figure 3 This is a flowchart of the size table steps of a sensor array security monitoring method for multi-target identification in one embodiment of this application;

[0050] Figure 4 This is a flowchart of a dynamic thresholding + binarization thresholding method for a sensor array security monitoring method for multi-target identification in one embodiment of this application;

[0051] Figure 5 This is a flowchart of a multi-target identification sensor array security monitoring method based on an embodiment of this application, which includes a water immersion thresholding method and a fixed thresholding method. Detailed Implementation

[0052] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0053] Furthermore, descriptions using terms such as "first" and "second" in this application are for descriptive purposes only (e.g., to distinguish identical or similar elements) and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, technical solutions from different embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If a combination of technical solutions is contradictory or impossible to implement, such a combination should be considered nonexistent and not within the scope of protection claimed in this application.

[0054] like Figures 1-3 As shown, this invention proposes a multi-target identification sensor array security monitoring method, comprising the following steps:

[0055] S1. Collect environmental data and target signal data within the monitoring area through a sensor array consisting of several sensors, wherein the sensors include a combination of at least two of acoustic sensors, vibration sensors, electromagnetic sensors, infrared sensors, and optical sensors;

[0056] Furthermore, in step S1, the sensor array is arranged according to a predetermined topology, which can be any one of grid, ring, or star topology, and the spacing between adjacent sensors is determined according to the area of ​​the monitoring area and the required monitoring accuracy.

[0057] For example, in open areas, a grid pattern can be used to achieve uniform coverage; a ring pattern may be more suitable for perimeter monitoring of specific areas; and a star pattern is advantageous when focused monitoring of the central area is required. The spacing between adjacent sensors is determined based on the area of ​​the monitoring region and the required monitoring accuracy to ensure that the coverage and resolution of the data acquisition meet the requirements. For example, for high-precision monitoring, the sensor spacing will be smaller;

[0058] S2. Perform data preprocessing on the collected environmental data and target signal data to eliminate noise and standardize the data format;

[0059] Specifically, in step S2, the denoising employs any one or more combinations of wavelet transform denoising, median filtering denoising, and Kalman filtering denoising; data preprocessing includes at least one of random cropping, flipping, rotating, and scaling.

[0060] S3. Perform feature extraction and fusion on the preprocessed sensor array data to generate a fused feature vector. For example, for acoustic sensor data, the root mean square and peak values ​​in the time domain, the spectral energy distribution in the frequency domain, and the Mel-frequency cepstral coefficients (MFCC) in the time and frequency domains can be extracted. For vibration sensor data, features such as vibration amplitude, frequency, and phase can be extracted. For array data, spatial distribution features, such as the angle of arrival (DOA), can also be extracted.

[0061] S4. Apply any one of the following thresholding methods to the fused feature vector: dynamic thresholding, fixed thresholding, binarization thresholding, and water immersion thresholding; and / or a combination of at least two thresholding methods; to extract potential target signals and generate preliminary target events;

[0062] Specifically, dynamic thresholding: The threshold is adaptively determined based on the gray-level distribution of local areas of the image, which is suitable for images with uneven lighting or complex backgrounds.

[0063] Fixed threshold: A global threshold is preset, which is suitable for scenes with uniform lighting and obvious contrast between defects and background.

[0064] Binarization threshold: Divides the gray values ​​of image pixels into two categories (usually 0 and 255), which is suitable for completely separating defects.

[0065] Watershed Transform: A morphology-based segmentation method suitable for separating adhered defects or extracting defects from complex backgrounds.

[0066] S5. Create a large table and a small table based on the preliminary target events. The large table is configured to store the complete context information of all preliminary target events.

[0067] Specifically, the large table stores the complete context information of the initial target event, including: associated event ID (used to uniquely identify an initial event), raw data index (pointing to raw sensor data, facilitating backtracking analysis), environmental context data (such as ambient temperature, humidity, light intensity, etc., which helps in subsequent analysis), fused feature vector (complete feature information), and event processing status marker (recording the current processing stage of the event).

[0068] The small table stores short-term, high-confidence preliminary target event summary information, including: event type identifier (preliminary event type), timestamp (event occurrence time), spatial stamp (event occurrence location), confidence score (event reliability assessment), and trigger sensor code (indicating which sensor(s) triggered the event).

[0069] S6. Using the high-confidence preliminary target event summary information in the small table as the core, and based on its timestamp, spatial stamp and event type, define a spatiotemporal window in the large table, and retrieve all preliminary target events within the window;

[0070] S7. Perform correlation analysis on the preliminary target events retrieved within the spatiotemporal window. Based on the correlation analysis results, merge several preliminary target events belonging to the same target object into a composite event, and retain the independent state of preliminary target events that are not belonging to the same target object, marking them as a set of events to be observed.

[0071] S8. Based on a preset multi-object classifier, target type identification and classification are performed on the merged composite events. The multi-object classifier is constructed based on a machine learning model or a deep learning model and outputs the type label and attribute information of each identified target. For example, convolutional neural network (CNN), recurrent neural network (RNN), support vector machine (SVM) or ensemble learning model can be used. Further, after processing the composite events, the classifier outputs the target type label, such as pedestrian, vehicle, animal, drone, etc., and attribute information, such as speed, direction, size, quantity, etc.

[0072] S9. Based on the identification results and the preset security policy library, a threat assessment is performed. If a real threat is determined, a predefined alarm mechanism is triggered. The alarm mechanism includes one or more of the following: audible and visual alarm, work order push, system log recording, and linkage with other security systems.

[0073] In this embodiment, a heterogeneous sensor array (acoustic / vibration / electromagnetic / infrared / optical combination) and adaptive topology layout are used to achieve full-dimensional coverage and dynamic adjustment of the monitoring area. The complementarity of data from different frequency bands and modalities significantly improves target resolution in complex scenes, especially maintaining high sensitivity in environments with changing lighting, occlusion interference, or mixed noise. Simultaneously, by combining multiple types of sensors and employing advanced feature extraction and fusion strategies (including feature stitching, PCA dimensionality reduction, and dynamic weighted fusion), the complementarity and redundancy of different sensor data can be fully utilized to generate more representative fused feature vectors, significantly improving the ability to perceive multiple targets in complex environments. Furthermore, various thresholding methods such as dynamic thresholding and binarized thresholding are introduced, especially the adaptive threshold calculated by combining historical statistical values, making the extraction of potential target signals more flexible and accurate, effectively reducing false alarms and false negatives, and improving the quality of initial target event generation.

[0074] On the other hand, a hierarchical storage mechanism using large and small tables enables refined management of initial target events. The small table is used for quick retrieval of high-confidence events, while the large table retains complete contextual information. Combined with correlation analysis within a spatiotemporal window, it can efficiently aggregate scattered initial events into composite events, laying a solid foundation for multi-target identification. Furthermore, based on a machine learning or deep learning-based multi-target classifier, it can accurately identify and classify the target type of composite events, outputting detailed target attribute information. Combined with a pre-set security policy library for threat assessment, it achieves intelligent management across the entire chain from perception and identification to early warning, significantly improving the response speed and decision-making capabilities of the security monitoring system.

[0075] Furthermore, each stage of the system, such as data acquisition, preprocessing, feature extraction, fusion, thresholding, and target recognition, adopts a modular design approach. This allows for the easy addition, deletion, or replacement of relevant modules based on specific monitoring scenarios and needs in practical applications. Thus, if a particular type of target in a specific area requires more precise monitoring, only targeted optimization or addition of the relevant feature extraction module is needed, without requiring large-scale modifications to the entire system.

[0076] For example, sensor arrays can be deployed around and in key areas within a residence to create a comprehensive security network. When an unauthorized intruder enters the monitoring range, the system can quickly identify them and issue an alarm. Simultaneously, the alarm information is pushed to the resident's mobile phone or other terminal devices via the network, enabling remote monitoring and real-time response.

[0077] By using a sensor array to monitor and analyze the power consumption of household appliances in real time, the energy consumption status of each device can be understood. Based on user habits and actual needs, the operating modes of the devices can be automatically adjusted to achieve energy conservation and emission reduction.

[0078] For families with elderly people living alone or children left behind by their parents, wearable sensors or sensor arrays installed in the living environment can monitor their health and activity patterns in real time. For example, when an elderly person falls or remains still for an extended period, the system will automatically trigger an alarm and notify family members or emergency personnel; for children, features such as electronic fences can be set up to prevent them from getting lost or entering dangerous areas.

[0079] In one embodiment, step S3, the feature extraction and fusion includes:

[0080] For the preprocessed data of each sensor node, features are extracted in parallel from multiple dimensions to form the initial feature vector of that node. The initial feature vector includes: time-domain features for describing the statistical characteristics and intensity changes of the signal on the time axis, frequency-domain features for revealing the frequency components of the signal and their distribution, time-frequency-domain features for simultaneously capturing the joint characteristics of the signal in time and frequency, and spatial distribution features for characterizing the source location and intensity distribution of the signal.

[0081] According to the preset fusion strategy, the initial feature vectors from different sensor nodes are integrated to generate the final fused feature vector; the fused feature vector includes a combination of at least two vectors among time domain features, frequency domain features, time-frequency domain features, and spatial distribution features.

[0082] The preset fusion strategy includes the following steps:

[0083] When the complementarity and consistency between features are higher than the preset first complementarity threshold, and the feature dimension is higher than the preset second dimension threshold, a feature splicing and fusion strategy is adopted to maximize the preservation of original information from multiple sensors.

[0084] When the redundancy between features exceeds the preset third redundancy threshold, or the feature dimension exceeds the preset fourth dimension threshold, and this state persists, it is determined that a principal component analysis dimensionality reduction and fusion strategy should be adopted to eliminate redundancy and reduce computational load.

[0085] If the conditions for either of the above two strategies are not triggered, the dynamic weighted fusion strategy will be adopted.

[0086] Furthermore, the first complementarity threshold is the minimum value of complementarity and consistency between features, used to determine whether feature splicing is applicable; the second dimension threshold is the minimum value of feature dimension, used in conjunction with the first complementarity threshold to determine whether feature splicing is applicable; the third redundancy threshold is the minimum value of redundancy between features, used to determine whether dimensionality reduction is triggered; and the fourth dimension threshold is the maximum value of feature dimension, used in conjunction with the third redundancy threshold to determine whether dimensionality reduction fusion is required.

[0087] Specifically, the feature concatenation and fusion strategy employs a method of directly concatenating initial feature vectors from different domains to form a high-dimensional fused feature vector. For example, it concatenates the temporal feature vector (mean, variance) describing the target's temporal dynamics with the spatial distribution feature vector (signal angle of arrival, intensity gradient) characterizing the target's azimuth distribution. By concatenating temporal features, frequency domain features, time-frequency domain features, and spatial distribution features, the most comprehensive feature representation is formed.

[0088] The principal component analysis (PCA) dimensionality reduction and fusion strategy concatenates several selected initial feature vectors, then applies PCA to extract principal components, forming a low-dimensional, decorrelated fused feature vector. For example, concatenating frequency domain feature vectors (spectral entropy) with time domain feature vectors (zero-crossing rate) for dimensionality reduction effectively extracts common information with high correlation between the frequency and time domains in vibration, sound, and other signals, eliminating redundancy, primarily used for equipment status monitoring. Concatenating spatial distribution feature vectors (signal strength values ​​from multiple sensors) with time domain feature vectors (short-time energy) for dimensionality reduction is used to extract the main features of the spatial-temporal pattern during target movement. Concatenating time domain, frequency domain, and spatial distribution feature vectors together and then performing PCA processing is used to extract the most representative comprehensive feature components from all available information.

[0089] The dynamic weighted fusion strategy treats feature vectors from different domains as different feature "channels" or "sequences," and dynamically calculates the weights of each feature domain through an attention mechanism to achieve weighted fusion. For example, the model learns the importance weights of time-domain feature sequences (statistics within a set of time windows) and time-frequency domain feature sequences (a set of wavelet coefficients). The model weights spatial distribution features (features from sensors in different orientations) and frequency domain features. The model simultaneously processes all feature domains—time, frequency, time-frequency, and spatial distribution—with the attention mechanism performing dual weighting at the feature level (calculating which type of feature is more important) and / or the sensor level (calculating which sensor's data is more reliable).

[0090] like Figure 4 As shown, in one embodiment, the thresholding method in step S4 employs a combination of dynamic thresholding and binarized thresholding, specifically including:

[0091] a- Dynamically calculates adaptive thresholds based on historical statistical values ​​of fused feature vectors;

[0092] b- Discretizes continuous feature vectors into a 0 / 1 state sequence by using a binarization threshold;

[0093] c- When the binarized state continues for more than the time window set by the dynamic threshold, it is determined to be a valid target event.

[0094] Specifically, adaptive threshold Based on feature vector history window The statistical measures within the mean are calculated, and common calculation methods include linear combinations of the mean and standard deviation:

[0095]

[0096] in, It represents the mean of a reference feature component (or the magnitude / mean of a feature vector) within a historical time window;

[0097] The standard deviation of the same reference characteristic component;

[0098] It is an adjustable parameter used to control the sensitivity of the threshold (e.g., α=2 or 3 corresponds to the 95% or 99.7% confidence interval of the normal distribution).

[0099] W is the length of the historical time window used for statistical calculations.

[0100] Furthermore, continuous feature vectors A certain scalar mapping For example: values ​​of a specific dimension, L2 norm, similarity to the reference template; and adaptive thresholds. Compare and generate a binary state sequence. :

[0101]

[0102] Define a duration counter Or use a length of Sliding time window; when binary state The duration is 1, exceeding the preset minimum duration. When a valid event occurs, it is determined that the event has occurred.

[0103] like Figure 5 As shown, in one embodiment, the thresholding method in step S4 employs a combination of immersion threshold and fixed threshold processing, specifically including:

[0104] d - Use a fixed threshold as the basic trigger condition;

[0105] e- The water immersion threshold method is used to perform regional growth analysis on the triggered signal;

[0106] f - When the coverage rate of the growth area reaches the preset ratio, it is determined to be a valid target event.

[0107] Specifically, let V be a two-dimensional feature map; for example, generated from sensor arrays, image pixels, or other spatial distribution features. Indicates position The eigenvalue at that location.

[0108] Use a fixed global threshold Perform preliminary binarization on the feature map to generate a seed point map. ;

[0109] ;

[0110] With each seed point Starting from this point, the region is expanded. The water immersion thresholding method typically uses a relatively low global threshold. And satisfy As a condition for growth.

[0111] Furthermore, a binary mask M is defined to mark the pixels that are ultimately submerged (belonging to the target region), initially... For each seed point, check its neighborhood (e.g., 4-connected or 8-connected neighborhood). pixels within :

[0112] and ;

[0113] The newly submerged points are added to the queue as new growth points, and this process is repeated until no new points can be submerged.

[0114] Calculate the area of ​​the grown region M and the entire Region of Interest (ROI). The ratio of area. When the coverage exceeds a preset ratio. When the event is determined to be a valid target event.

[0115] This invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned multi-target recognition sensor array security monitoring method. When executed, the computer program may include the flow of the respective embodiments of the multi-target recognition sensor array security monitoring method described above. Any references to memory, storage, databases, or other media used in this application and in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0116] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, apparatus, article, or sensor array security monitoring method for multi-target identification that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or sensor array security monitoring method for multi-target identification. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or sensor array security monitoring method for multi-target identification that includes that element.

[0117] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A security monitoring method for a sensor array with multi-target recognition, characterized in that, Includes the following steps: S1. A sensor array consisting of at least two different types of sensors collects environmental data and target signal data within the monitoring area. The sensor types include acoustic sensors, vibration sensors, electromagnetic sensors, infrared sensors, and optical sensors. The sensor array is arranged according to a predetermined grid, ring, or star topology. The spacing between adjacent sensors is determined based on the area of ​​the monitoring area and the required monitoring accuracy. S2. Perform data preprocessing on the collected environmental data and target signal data to eliminate noise and standardize the data format; S3. Perform feature extraction and fusion on the preprocessed sensor array data to generate a fused feature vector; S4. Apply a thresholding method to the fused feature vector, which is any one or a combination of at least two of the following: dynamic threshold, fixed threshold, binarization threshold, and water immersion threshold, to extract potential target signals and generate preliminary target events. S5. Create and store a large table and a small table based on the preliminary target event, wherein the large table stores the complete context information of the preliminary target event, and the small table stores short-term high-confidence summary information of the preliminary target event; S6. Using the high-confidence event summary information in the small table as the core, and based on the timestamp, spatial stamp, and event type included therein, define a spatiotemporal window in the large table, and retrieve all preliminary target events within the window; S7. Perform correlation analysis on the preliminary target events retrieved within the spatiotemporal window, and merge several preliminary target events belonging to the same target object into a composite event based on the analysis results, and mark the preliminary target events that do not belong to the same target object as a set of events to be observed. S8. Based on a preset multi-target classifier, identify and classify the target type of the composite event, and output the type label and attribute information of each identified target; S9. Based on the identification results and the preset security policy library, a threat assessment is performed. If a real threat is determined, a predefined alarm mechanism is triggered.

2. The multi-target identification sensor array security monitoring method according to claim 1, characterized in that, In step S2, the denoising in the data preprocessing adopts any one or more combinations of wavelet transform denoising, median filtering denoising, and Kalman filtering denoising; the data preprocessing also includes at least one operation of random cropping, flipping, rotating, and scaling.

3. The multi-target identification sensor array security monitoring method according to claim 1, characterized in that, In step S3, the feature extraction and fusion includes: For the preprocessed data of each sensor node, features are extracted from multiple dimensions, including time domain, frequency domain, time-frequency domain, and spatial distribution, to form the initial feature vector of that node; According to the preset fusion strategy, the initial feature vectors from different sensor nodes are integrated to generate the final fused feature vector. The fused feature vector includes a combination of at least two vectors among time domain features, frequency domain features, time-frequency domain features, and spatial distribution features.

4. The multi-target identification sensor array security monitoring method according to claim 3, characterized in that, The preset fusion strategy includes: When the complementarity and consistency between features are higher than the preset first complementarity threshold, and the feature dimension is higher than the preset second dimension threshold, a feature splicing and fusion strategy is adopted. When the redundancy between features exceeds the preset third redundancy threshold, or the feature dimension exceeds the preset fourth dimension threshold, a principal component analysis dimensionality reduction and fusion strategy is adopted. If the conditions for either of the above two strategies are not triggered, a dynamic weighted fusion strategy will be adopted.

5. The multi-target identification sensor array security monitoring method according to claim 1, characterized in that, In step S4, the thresholding method employs a combination of dynamic thresholding and binarized thresholding, specifically including: Dynamically calculate adaptive thresholds based on historical statistical values ​​of fused feature vectors; The continuous feature vector is discretized into a 0 / 1 state sequence by using a binarization threshold; When the binarized state continues for more than the time window set by the dynamic threshold, it is determined to be a valid target event.

6. The multi-target identification sensor array security monitoring method according to claim 5, characterized in that, The formula for calculating the adaptive threshold is: ； Where μ represents the mean of a reference feature component within the historical time window, σ represents the standard deviation of the same reference feature component, and α is an adjustable sensitivity parameter.

7. The multi-target identification sensor array security monitoring method according to claim 1, characterized in that, In step S4, the thresholding method employs a combination of immersion threshold and fixed threshold processing, specifically including: A fixed threshold is used as the basic trigger condition; The water immersion threshold method was used to perform regional growth analysis on the triggered signal; When the coverage rate of the growth area reaches the preset ratio, it is determined to be a valid target event.

8. The multi-target identification sensor array security monitoring method according to claim 1, characterized in that, In step S5: The information stored in the large table includes: associated event ID, original data index, environmental context data, fused feature vector, and event processing status marker; The information stored in the small table includes: event type identifier, timestamp, spatial stamp, confidence score, and trigger sensor code.

9. A sensor array security monitoring method for multi-target identification according to claim 1, characterized in that, The multi-objective classifier is built based on a machine learning model or a deep learning model, including one of a convolutional neural network, a recurrent neural network, a support vector machine, or an ensemble learning model.

10. A sensor array security monitoring method for multi-target identification according to claim 1, characterized in that, In step S9, the alarm mechanism includes one or more of the following: audible and visual alarm, work order push, system log recording, and linkage with other security systems.

Images