Perimeter intrusion monitoring system based on cooperation of infrared thermal imaging and intelligent algorithm

Abstract

The invention relates to the technical field of perimeter intrusion monitoring, in particular to a perimeter intrusion monitoring system based on cooperation of infrared thermal imaging and an intelligent algorithm. The perimeter intrusion monitoring system based on infrared thermal imaging and intelligent algorithm cooperation comprises a multi-source sensing layer, an edge computing node, a cloud analysis platform and a response execution layer. According to the perimeter intrusion monitoring system based on cooperation of infrared thermal imaging and an intelligent algorithm, environmental noise interference is effectively suppressed through a dynamic threshold compensation mechanism, and accurate classification of personnel, vehicles, unmanned aerial vehicles and animals is realized by combining multi-modal feature fusion of infrared thermal imaging, visible light vision and millimeter wave radar; according to the method, intention modes of passing, observation, invasion and the like can be identified, threat level dynamic adaptation is realized in combination with a hierarchical response system, the technical bottlenecks of poor environmental adaptability, high target misjudgment rate, behavior analysis deficiency and the like of a traditional system are solved, and the intelligent level and all-weather protection capability of perimeter security and protection are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of perimeter intrusion detection technology, and in particular to a perimeter intrusion detection system based on the synergy of infrared thermal imaging and intelligent algorithms. Background Technology

[0002] Perimeter security systems are a crucial technological means of ensuring the security of critical infrastructure such as airports, oil and gas pipelines, and military bases. Their performance directly impacts security protection levels and emergency response efficiency. Traditional infrared perimeter security systems primarily rely on passive infrared detectors or fixed-threshold thermal imaging equipment, detecting changes in ambient temperature to determine the presence of intrusion. However, with the diversification of intrusion methods and the increasing complexity of environments, traditional systems have revealed numerous technical bottlenecks in practical applications, necessitating intelligent upgrades and system reconstruction.

[0003] I. Insufficient environmental adaptability, resulting in a persistently high false alarm rate.

[0004] Traditional infrared perimeter security systems generally use fixed detection thresholds, failing to adequately consider the impact of environmental changes on thermal imaging signals. For example, in heavy rain, the absorption and scattering of infrared radiation by raindrops significantly alters the thermal signal distribution, causing the system to misidentify intrusion events, with a false alarm rate as high as 68% (field test data from an oil field in 2024). Furthermore, in hot weather, the increased thermal radiation from high-heat-capacity materials such as asphalt can cause drastic fluctuations in background noise in thermal images, resulting in a false alarm rate as high as 63%. These dynamic changes in environmental factors are not effectively modeled and compensated for, severely limiting the system's stability and reliability.

[0005] II. Limited target recognition capabilities and low classification accuracy.

[0006] Existing systems primarily rely on a single temperature threshold for target identification, lacking multi-dimensional perception and analysis of target features. For example, human heat sources typically appear as clear heat spots around 37°C, while animal heat sources often present as blurry heat areas below 35°C. Drones, due to their high-speed rotor rotation, exhibit dynamic features of alternating hot and cold temperatures in infrared images. However, traditional systems cannot effectively extract this feature information, making it difficult to accurately distinguish between different targets such as people, animals, and drones, with a false positive rate generally exceeding 50%. These technical deficiencies not only reduce the system's practicality but also increase the handling burden on security personnel.

[0007] Third, a lack of behavioral analysis capabilities makes it difficult to predict intentions.

[0008] Current perimeter security systems mostly focus on detecting whether an intrusion has occurred, lacking continuous tracking of target movement trajectories and intelligent analysis of behavioral patterns. For example, these systems cannot effectively distinguish between different behavioral intentions such as "passing by" and "deliberate sabotage," leading to many non-threatening events being misjudged as emergency intrusions, resulting in wasted security resources and inefficient response mechanisms. Predicting the intent behind intrusion behavior is key to improving system intelligence; however, traditional systems struggle to meet this requirement due to limitations such as single data collection dimensions and outdated algorithm models.

[0009] In summary, traditional infrared perimeter security systems face the following critical technical challenges that urgently need to be addressed:

[0010] 1. Poor environmental adaptability: When facing complex and ever-changing natural environments (such as rainstorms, high temperatures, and different surface materials), the system suffers from a high false alarm rate due to the lack of a dynamic parameter adjustment mechanism.

[0011] 2. Low target recognition accuracy: Relying solely on single temperature information for target classification makes it difficult to distinguish between different heat source targets such as people, animals, and drones, resulting in frequent system misjudgments;

[0012] 3. Lack of behavioral intent analysis capability: The system is unable to continuously track the target's movement trajectory and model behavioral patterns, and cannot upgrade from "whether there is an intrusion" to "predicting intrusion intent".

[0013] To address these issues, a perimeter intrusion detection system based on the synergy of infrared thermal imaging and intelligent algorithms is provided. Summary of the Invention

[0014] This invention provides a perimeter intrusion detection system based on the collaboration of infrared thermal imaging and intelligent algorithms. It aims to break through the technical bottlenecks of traditional systems and significantly improve the intelligence level and environmental adaptability of perimeter security systems by introducing key technologies such as dynamic threshold generation module, multimodal target classifier, and target behavior analysis.

[0015] To achieve the above objectives, the present invention adopts the following technical solution:

[0016] A perimeter intrusion detection system based on the collaboration of infrared thermal imaging and intelligent algorithms includes a multi-source perception layer, edge computing nodes, a cloud analysis platform, and a response execution layer.

[0017] The multi-source sensing layer is equipped with an infrared thermal imager, a visible light camera, a millimeter-wave radar, and a meteorological sensor to simultaneously acquire multi-dimensional data;

[0018] The edge computing node is equipped with a dynamic threshold generation module, which calculates compensation factors in real time through the multi-dimensional data to generate dynamic thresholds.

[0019] The cloud analytics platform is equipped with a multimodal target classifier and an LSTM behavior analyzer. The multimodal target classifier is used to classify targets, and the LSTM behavior analyzer is used to predict the intent of target behavior.

[0020] The response execution layer triggers a tiered response based on the analysis results from the cloud analytics platform.

[0021] As a further description of the above technical solution:

[0022] The multi-dimensional data includes the temperature difference matrix of the target captured by the infrared thermal imager, the RGB image of the target captured by the visible light camera, the motion trajectory of the target captured by the millimeter-wave radar, and the ambient temperature, humidity, solar radiation flux, wind speed, and precipitation rate captured by the meteorological sensor.

[0023] As a further description of the above technical solution:

[0024] The multimodal target classifier extracts the infrared thermal features, visible light morphological features, and radar motion features of the target through the multidimensional data. The multimodal target classifier classifies the target based on the infrared thermal features, visible light morphological features, and radar motion features, combined with the dynamic threshold. The target categories are divided into personnel, vehicles, drones, and animals.

[0025] As a further description of the above technical solution:

[0026] The LSTM behavior analyzer is used to predict the intent of target behavior. The output results are divided into passing by, observing, and intruding.

[0027] As a further description of the above technical solution:

[0028] The hierarchical response of the response execution layer includes audible and visual alarm prompts, drone tracking, and security dispatch.

[0029] As a further description of the above technical solution:

[0030] The final dynamic threshold T of the dynamic threshold generation module d Based on the benchmark threshold T b The compensation factor comp and the wind speed correction term are jointly determined, and the formula is:

[0031] T d =T b ·comp·(1+0.01·max(0,v w -5))

[0032] Where: v w The ambient wind speed is captured by the meteorological sensor.

[0033] Wind speed correction term: wind speed v w When the speed is >5m / s, wind vibration will cause fluctuations in the infrared signal, and the threshold needs to be linearly amplified to suppress noise.

[0034] Baseline threshold T b Based on the preset protection level, standard level T b =2.5, Advanced Level T b =1.8, Special Grade T b =1.2;

[0035] The compensation factor comp is divided into rainy day compensation factor comp1, thermal accumulation compensation factor comp2 and normal environment compensation factor comp3 according to the scenario.

[0036] The formula for the rain compensation factor comp1 is:

[0037] comp1 = 1.2 + 0.05·R

[0038] Where: R is the precipitation rate captured by the meteorological sensor, in mm / h; the rain compensation factor trigger condition is: the meteorological sensor detects a precipitation rate R>0;

[0039] The formula for the thermal accumulation compensation factor comp2 is:

[0040]

[0041] Where: T is the ambient temperature captured by the meteorological sensor, in degrees Celsius; the triggering condition for the thermal accumulation compensation factor is: the surface material G = ASPHALT (asphalt) and the solar radiation flux S captured by the meteorological sensor > 800 W / m;

[0042] The conventional environmental compensation factor comp3 is obtained by looking up historical experience data in a table, as shown in the following table:

[0043] Weather type Time period <![CDATA[Compensation factor comp3]]> sunny Dawn (4-6 hours) 1.45 sunny Daytime (7-18 hours) 1.35 sunny Dusk (7-8 pm) 1.25 sunny Nighttime (21-3h) 1.15 partly cloudy Dawn (4-6 hours) 1.38 partly cloudy Daytime (7-18 hours) 1.28 partly cloudy Dusk (7-8 pm) 1.2 partly cloudy Nighttime (21-3h) 1.1 cloudy day Dawn (4-6 hours) 1.32 cloudy day Daytime (7-18 hours) 1.22 cloudy day Dusk (7-8 pm) 1.18 cloudy day Nighttime (21-3h) 1.08

[0044] Among them, the triggering condition for the conventional environmental compensation factor is a non-rainy day and a non-thermal accumulation environment.

[0045] As a further description of the above technical solution:

[0046] The classification method of the multimodal target classifier includes the following steps:

[0047] Step 1: Detect the thermal radiation intensity of the target surface using the infrared thermal imager and extract the following key features:

[0048] Hot spot distribution histogram: The target hot spot energy is divided into 10 intervals in the spatial domain, and the energy proportion of each interval is statistically analyzed to describe the spatial distribution uniformity of the hot spot.

[0049] Maximum temperature difference: Calculates the difference between the highest and lowest temperatures on the target surface (unit: °C), reflecting the degree of concentration of heat sources inside the target;

[0050] Thermal radiation area percentage: The ratio of the area covered by the hot spot to the total projected area of ​​the target (unit: %), used to distinguish the relative size of the heat-generating area;

[0051] The visible light camera extracts the morphological features of the target using target detection and image analysis algorithms.

[0052] YOLOv8 contour complexity: Used to quantify the complexity of a target structure, calculated based on the frequency of curvature changes in the target edge contour, using the following formula:

[0053]

[0054] Where: L is the total length of the profile, θ i Let be the tangent angle at the i-th edge point;

[0055] Aspect ratio: The ratio of the target's height to its width, reflecting the target's geometric shape;

[0056] The millimeter-wave radar extracts the kinematic features of the target through the Doppler effect and target tracking algorithms:

[0057] Trajectory curvature: The degree of curvature of the target motion trajectory, defined as the rate of change of the tangent direction of the path, used to distinguish between linear motion and directional motion;

[0058] Speed ​​range: The instantaneous speed of the target's movement (unit: m / s), reflecting the target's motion capability;

[0059] Step 2: Initial screening based on radar motion characteristics, coarse classification based on dynamic constraints: speed <1m / s and trajectory curvature >0.5rad / m: initially identified as "animal" candidate set;

[0060] Speed ​​≤ 1 m / s ≤ 5 m / s and trajectory curvature ≤ 0.3 rad / m: Preliminary assessment as a candidate set for "personnel / drone".

[0061] Speed ​​≥ 5 m / s and trajectory curvature ≤ 0.3 rad / m: Preliminarily identified as a candidate set for "vehicle";

[0062] Step 3: Infrared thermal property re-screening, fine classification based on thermodynamic constraints.

[0063] Candidate set for "animals": If the maximum temperature difference is <5℃, the "animal" category is pre-confirmed;

[0064] Candidate set of "personnel / drone swarms": If the maximum temperature difference is ≥5℃ and the hot spot distribution histogram is ≈30%-50%, it is initially determined to be a candidate set of "personnel" and will proceed to the next step of visible light morphology verification; If the maximum temperature difference is ≥5℃ and the hot spot distribution histogram is ≥60%, it is initially determined to be a candidate set of "drone swarms" and will proceed to the next step of visible light morphology verification.

[0065] Candidate set of "vehicles"; if the maximum temperature difference is ≥5℃ and the hot spot distribution histogram is <30%, the "vehicle" category is pre-confirmed.

[0066] Step 4: Visible light morphology verification, final confirmation based on geometric features.

[0067] Candidate set of "personnel": moderate profile complexity, confirming the "personnel" category;

[0068] Candidate set for "drone": Aspect ratio ≈ 0.3, confirming the "drone" category;

[0069] Candidate set for "Vehicle": The outline complexity is the highest, and the aspect ratio is >1.0, confirming the "Vehicle" category;

[0070] The candidate set for "animals" has the lowest outline complexity and no fixed aspect ratio, thus confirming the "animal" category.

[0071] As a further description of the above technical solution:

[0072] The LSTM behavior analyzer takes the target's motion trajectory in the spatiotemporal dimension as input data, and the specific processing flow includes the following steps:

[0073] Step 1: Data Input and Preprocessing

[0074] Data Acquisition: Obtain the target's three-dimensional coordinate sequence {(x1,y1,t1),(x2,y2,t2),...,(x...} using millimeter-wave radar. n ,y n ,t n The sampling frequency is 10Hz.

[0075] Data cleaning: outliers are removed, and missing values ​​in the time window are filled using linear interpolation;

[0076] Normalization: Mapping coordinate values ​​to an interval, the formula is:

[0077]

[0078] Ensure that data of different dimensions are trained on the same scale;

[0079] Window partitioning: The sequence is divided according to the time step Δt = 1 second, and a sliding window is constructed to form a supervised learning dataset;

[0080] Step 2: Feature Extraction

[0081] Key temporal features are extracted from the original trajectory. These key temporal features include basic features and higher-order features. The basic features include velocity, acceleration, and trajectory curvature. The higher-order features include dwell index and path complexity.

[0082] The dwell index SI, representing the percentage of time the target remains stationary, is calculated using the following formula:

[0083]

[0084] Where: δ is an indicator function, which takes the value 1 when the condition is met and 0 otherwise; the threshold of 2 seconds is used to distinguish between a brief stop and passing by.

[0085] The path complexity PC is the ratio of the actual path length to the straight-line distance, and its formula is:

[0086]

[0087] Where: PC value range is [0,1], the larger the value, the more circuitous the path, which may be an observation or exploratory behavior;

[0088] Step 3: Network Architecture and Computation Process

[0089] The LSTM behavior analyzer employs a bidirectional long short-term memory network combined with an attention mechanism to capture temporal dependencies and focus on key behavioral segments. Its specific structure is as follows:

[0090] Input layer: Receives the key temporal features obtained in the second step;

[0091] Bidirectional LSTM layer:

[0092] Forward LSTM: Processes the sequence in chronological order and outputs the hidden state.

[0093] Reverse LSTM: Processes the sequence in reverse order and outputs the hidden state.

[0094] Feature fusion: concatenating the bidirectional hidden states into Capture global time-series dependencies;

[0095] Attention mechanism:

[0096] Calculate the attention weight α at each time step t The formula is:

[0097]

[0098] Among them: W qA learnable query matrix used to measure the importance of time steps;

[0099] Generate context vectors Focus on key behavioral time steps;

[0100] Fully connected layer: Maps the context vector c to a 3D output space, and obtains the behavior probability distribution through the Softmax function.

[0101] P = Softmax(W o ·c+b o )

[0102] Among them: W o To output the weight matrix, b o As a bias term, the output P = {p 路过 p 观察 p 入侵 The sum of probabilities is 1.

[0103] Step 4: Output and Decision Logic

[0104] Single-frame decision: If p at a certain time step 入侵 If the value is >0.6, it is marked as "suspected intrusion" and a primary warning (audio and visual alarm) is triggered.

[0105] Timing verification: If p for 3 consecutive time steps 入侵 If the value is greater than 0.7, it is considered an "intrusion confirmed" and an advanced response (drone tracking, security dispatch) is triggered.

[0106] Intentional grading: If p 观察 A value >0.5 is considered "malicious observation". 路过 A value >0.5 is considered "harmless passing by," assisting security personnel in their decision-making.

[0107] The present invention has the following beneficial effects:

[0108] The perimeter intrusion monitoring system based on the collaboration of infrared thermal imaging and intelligent algorithms described in this invention effectively suppresses environmental noise interference through a dynamic threshold compensation mechanism. By combining multimodal feature fusion of infrared thermal imaging, visible light vision, and millimeter-wave radar, it achieves accurate classification of personnel, vehicles, drones, and animals. The behavior spatiotemporal analysis model based on LSTM network can identify intent patterns such as "passing by, observing, and intruding," and combined with a hierarchical response system, it achieves dynamic adaptation to threat levels. This solves the technical bottlenecks of traditional systems, such as poor environmental adaptability, high target misjudgment rate, and lack of behavior analysis, and significantly improves the intelligence level and all-weather protection capability of perimeter security. Attached Figure Description

[0109] Figure 1This is a schematic diagram illustrating the architecture of the perimeter intrusion detection system based on the collaboration of infrared thermal imaging and intelligent algorithms described in this invention.

[0110] Figure 2 This is a schematic diagram of the dynamic threshold generation process described in this invention;

[0111] Figure 3 This is a schematic diagram of the LSTM behavior analyzer described in this invention; Detailed Implementation

[0112] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0113] A perimeter intrusion detection system based on the collaboration of infrared thermal imaging and intelligent algorithms includes a multi-source perception layer, edge computing nodes, a cloud analysis platform, and a response execution layer.

[0114] The multi-source sensing layer is equipped with an infrared thermal imager, a visible light camera, a millimeter-wave radar, and a meteorological sensor to simultaneously acquire multi-dimensional data;

[0115] The edge computing node is equipped with a dynamic threshold generation module, which calculates compensation factors in real time through the multi-dimensional data to generate dynamic thresholds.

[0116] The cloud analytics platform is equipped with a multimodal target classifier and an LSTM behavior analyzer. The multimodal target classifier is used to classify targets, and the LSTM behavior analyzer is used to predict the intent of target behavior.

[0117] The response execution layer triggers a tiered response based on the analysis results from the cloud analytics platform.

[0118] As a further description of the above technical solution:

[0119] The multi-dimensional data includes the temperature difference matrix of the target captured by the infrared thermal imager, the RGB image of the target captured by the visible light camera, the motion trajectory of the target captured by the millimeter-wave radar, and the ambient temperature, humidity, solar radiation flux, wind speed, and precipitation rate captured by the meteorological sensor.

[0120] As a further description of the above technical solution:

[0121] The multimodal target classifier extracts the infrared thermal features, visible light morphological features, and radar motion features of the target through the multidimensional data. The multimodal target classifier classifies the target based on the infrared thermal features, visible light morphological features, and radar motion features, combined with the dynamic threshold. The target categories are divided into personnel, vehicles, drones, and animals.

[0122] As a further description of the above technical solution:

[0123] The LSTM behavior analyzer is used to predict the intent of target behavior. The output results are divided into passing by, observing, and intruding.

[0124] As a further description of the above technical solution:

[0125] The hierarchical response of the response execution layer includes audible and visual alarm prompts, drone tracking, and security dispatch.

[0126] As a further description of the above technical solution:

[0127] The final dynamic threshold T of the dynamic threshold generation module d Based on the benchmark threshold T b The compensation factor comp and the wind speed correction term are jointly determined, and the formula is:

[0128] T d =T b ·comp·(1+0.01·max(0,v w -5))

[0129] Where: v w The ambient wind speed is captured by the meteorological sensor.

[0130] Wind speed correction term: wind speed v w At speeds >5 m / s, wind vibrations can cause fluctuations in the infrared signal, requiring linear amplification of the threshold to suppress noise (e.g., wind speed v). w When the speed is 10 m / s, the correction term is 1 + 0.01 × (10 - 5) = 1.05.

[0131] Baseline threshold T b Based on the preset protection level, standard level T b =2.5, Advanced Level T b =1.8, Special Grade T b =1.2;

[0132] The compensation factor comp is divided into rainy day compensation factor comp1, thermal accumulation compensation factor comp2 and normal environment compensation factor comp3 according to the scenario.

[0133] The formula for the rain compensation factor comp1 is:

[0134] comp1 = 1.2 + 0.05·R

[0135] Where: R is the precipitation rate captured by the meteorological sensor, in mm / h; the rain compensation factor trigger condition is: the meteorological sensor detects a precipitation rate R>0 (i.e., it is raining), for example: during heavy rain (R=50mm / h), comp1=1.2+0.05×50=3.7;

[0136] The formula for the thermal accumulation compensation factor comp2 is:

[0137]

[0138] Where: T is the ambient temperature captured by the meteorological sensor, in degrees Celsius; the trigger condition for the thermal accumulation compensation factor is: the surface material G = ASPHALT (asphalt) and the solar radiation flux S captured by the meteorological sensor > 800 W / m (strong sunlight), for example: in summer high temperature (T = 40℃), comp2 = 1.8 - 40 / 40 = 0.8;

[0139] The conventional environmental compensation factor comp3 is obtained by looking up historical experience data in a table, as shown in the following table:

[0140]

[0141]

[0142] Among them, the triggering condition for the conventional environmental compensation factor is a non-rainy day and a non-thermal accumulation environment.

[0143] As a further description of the above technical solution:

[0144] The classification method of the multimodal target classifier includes the following steps:

[0145] Step 1: Detect the thermal radiation intensity of the target surface using the infrared thermal imager and extract the following key features:

[0146] Hot spot distribution histogram: The target hot spot energy is divided into 10 intervals in the spatial domain, and the energy proportion of each interval is statistically analyzed to describe the spatial distribution uniformity of the hot spot (e.g., the hot spot of a rotary-wing UAV shows a periodic distribution due to the rotation of the rotor, and the histogram kurtosis is low; the hot spot of a person is concentrated in the torso area, and the kurtosis is high).

[0147] Maximum temperature difference: Calculate the difference between the highest and lowest temperatures on the target surface (unit: °C), reflecting the degree of heat concentration inside the target (for personnel, due to heat generation from torso metabolism, the maximum temperature difference is ≥5 °C; for vehicles, due to engine heat dissipation, the maximum temperature difference is ≥5 °C; for animals, due to uniform heat diffusion from fur coverage, the maximum temperature difference is <5 °C).

[0148] Thermal radiation area percentage: The ratio of the area covered by the hot spot to the total projected area of ​​the target (unit: %), used to distinguish the relative size of the heat-generating area (when the rotor motor of the UAV is rotating at high speed, the thermal radiation area percentage is ≥60%; the thermal radiation area percentage of the human torso is about 30-50%; the thermal radiation area percentage of the vehicle engine area is <30%).

[0149] The visible light camera extracts the morphological features of the target using target detection and image analysis algorithms.

[0150] YOLOv8 contour complexity: calculated based on the curvature change frequency of the target edge contour, using the following formula:

[0151]

[0152] Where: L is the total length of the profile, θ i Let be the tangent angle at the i-th edge point;

[0153] Used to quantify the complexity of the target structure (vehicles have details such as wheels and headlights, so the YOLOv8 contour complexity is: vehicles > people > animals).

[0154] Aspect ratio: The ratio of the target's height to its width, reflecting the target's geometric shape (Drones have a width-to-height ratio of ≈0.3 due to their flat and wide design; personnel have a width-to-height ratio of ≈0.8-1.2 due to their humanoid structure; vehicles have a width-to-height ratio of >1.0 because their height is greater than their width).

[0155] The millimeter-wave radar extracts the kinematic features of the target through the Doppler effect and target tracking algorithms:

[0156] Trajectory curvature: The degree of curvature of the target's trajectory, defined as the rate of change of the tangent direction of the path, used to distinguish between linear motion and directional motion (trajectory curvature ≈ 0 when a vehicle is moving at a constant speed; trajectory curvature ≈ 0.1~0.3 rad / m when a person is walking; trajectory curvature > 0.5 rad / m when an animal is moving randomly);

[0157] Speed ​​range: The instantaneous speed of the target's movement (unit: m / s), reflecting the target's mobility (human walking speed ≈ 0.8~1.8m / s; vehicle speed ≥ 5m / s; drone hovering or low-speed cruising ≈ 0.5~3m / s);

[0158] Step 2: Initial screening based on radar motion characteristics, coarse classification based on dynamic constraints: speed <1m / s and trajectory curvature >0.5rad / m: initially identified as "animal" candidate set;

[0159] Speed ​​≤ 1 m / s ≤ 5 m / s and trajectory curvature ≤ 0.3 rad / m: Preliminary assessment as a candidate set for "personnel / drone".

[0160] Speed ​​≥ 5 m / s and trajectory curvature ≤ 0.3 rad / m: Preliminarily identified as a candidate set for "vehicle";

[0161] Step 3: Infrared thermal property re-screening, fine classification based on thermodynamic constraints.

[0162] Candidate set of "animals": If the maximum temperature difference is <5℃ (animals have small body temperature differences due to their fur coverage and uniform metabolism), the "animal" category will be pre-confirmed.

[0163] Candidate sets for "personnel / drone swarms": those with a maximum temperature difference ≥ 5℃ (heat generation from trunk metabolism) and a hot spot distribution histogram ≈ 30%-50% (hot spots concentrated on the trunk) are initially identified as "personnel" candidates and will proceed to the next step of visible light morphology verification; those with a maximum temperature difference ≥ 5℃ (heat generation from rotor motors) and a hot spot distribution histogram ≥ 60% (hot spots covering the rotor area) are initially identified as "drone swarms" candidates and will proceed to the next step of visible light morphology verification.

[0164] Candidate set of "vehicles"; maximum temperature difference ≥5℃ and hot spot distribution histogram <30% (vehicles generate concentrated heat from the engine, but the engine area accounts for a small proportion), pre-confirm the "vehicle" category;

[0165] Step 4: Visible light morphology verification, final confirmation based on geometric features.

[0166] Candidate set of "personnel": moderate outline complexity (no structural details of vehicles, more streamlined than animals), aspect ratio ≈ 0.8-1.2 (consistent with humanoid structure), confirming the "personnel" category;

[0167] Candidate set of "drones": aspect ratio ≈ 0.3 (significantly lower than that of people and vehicles, with a flat and wide outline and no complex structure), confirming the "drone" category;

[0168] Candidate set for "vehicle": The outline complexity is the highest (including details such as wheels and headlights), and the aspect ratio is >1.0 (meets the geometric feature of "height > width"), confirming the "vehicle" category;

[0169] Candidate set for "animals": Lowest outline complexity (quadruped or streamlined), no fixed aspect ratio, confirming the "animal" category.

[0170] As a further description of the above technical solution:

[0171] The LSTM behavior analyzer takes the target's motion trajectory in the spatiotemporal dimension as input data, and the specific processing flow includes the following steps:

[0172] Step 1: Data Input and Preprocessing

[0173] Data Acquisition: Obtain the target's three-dimensional coordinate sequence {(x1, y1, t1), {x2, y2, t2), ..., (x...} using millimeter-wave radar. n y n , t n The sampling frequency is 10Hz.

[0174] Data cleaning: Remove outliers (such as noisy data with instantaneous velocity changes > 5 m / s), and use linear interpolation to fill in missing values ​​in the time window;

[0175] Normalization: Mapping coordinate values ​​to an interval, the formula is:

[0176]

[0177] Ensure that data of different dimensions are trained on the same scale;

[0178] Window partitioning: The sequence is partitioned according to the time step Δt = 1 second, and a sliding window is constructed (window size = 30 seconds, step size = 10 seconds) to form a supervised learning dataset;

[0179] Step 2: Feature Extraction

[0180] Key temporal features are extracted from the original trajectory. These key temporal features include basic features and higher-order features. The basic features include velocity, acceleration, and trajectory curvature. The higher-order features include dwell index and path complexity.

[0181] The dwell index SI, representing the percentage of time the target remains stationary, is calculated using the following formula:

[0182]

[0183] Where: δ is an indicator function, which takes the value 1 when the condition is met and 0 otherwise; the threshold of 2 seconds is used to distinguish between a brief stop and passing by.

[0184] The path complexity PC is the ratio of the actual path length to the straight-line distance, and its formula is:

[0185]

[0186] Where: PC value range is [0,1], the larger the value, the more circuitous the path, which may be an observation or exploratory behavior;

[0187] Step 3: Network Architecture and Computation Process

[0188] The LSTM behavior analyzer employs a bidirectional long short-term memory network (Bi-LSTM) combined with an attention mechanism to capture temporal dependencies and focus on key behavioral segments. Its specific structure is as follows:

[0189] Input layer: Receives the key temporal features obtained in the second step;

[0190] Bidirectional LSTM layer:

[0191] Forward LSTM: Processes the sequence in chronological order and outputs the hidden state.

[0192] Reverse LSTM: Processes the sequence in reverse order and outputs the hidden state.

[0193] Feature fusion: concatenating the bidirectional hidden states into Capture global time-series dependencies;

[0194] Attention mechanism:

[0195] Calculate the attention weight α at each time step t The formula is:

[0196]

[0197] Among them: W q A learnable query matrix used to measure the importance of time steps;

[0198] Generate context vectors Focus on key behavioral time steps (such as stopping, turning, etc.);

[0199] Fully connected layer: Maps the context vector c to a 3D output space, and obtains the behavior probability distribution through the Softmax function.

[0200] P = Softmax(W o ·c+b o )

[0201] Among them: W o To output the weight matrix, b o As a bias term, the output P = {p 路过 p 观察 p 入侵 (The sum of probabilities is 1);

[0202] Step 4: Output and Decision Logic

[0203] Single-frame decision: If p at a certain time step 入侵 If the value is >0.6, it is marked as "suspected intrusion" and a primary warning (audio and visual alarm) is triggered.

[0204] Timing verification: If p for 3 consecutive time steps 入侵 If the value is greater than 0.7, it is considered an "intrusion confirmed" and an advanced response (drone tracking, security dispatch) is triggered.

[0205] Intentional grading: If p观察 A value >0.5 is considered "malicious observation". 路过 A value >0.5 is considered "harmless passing by," assisting security personnel in their decision-making.

[0206] Example

[0207] Scenario Background: The perimeter protection system of a national-level nuclear power plant is deployed in a humid and windy coastal area, with the protection level set to "High" (baseline threshold T). b =1.8). At 14:00 on a certain day (during a sunny daytime period), the system detected an abnormal heat source approaching the north fence.

[0208] 1. Data acquisition in the multi-source sensing layer

[0209] Infrared thermal imager: captures the target temperature difference matrix, showing that hot spots are concentrated in the torso area (maximum temperature difference 6.2℃), with a thermal radiation area accounting for 41%.

[0210] Visible light camera: The detected target outline has an aspect ratio of 0.95 and a moderate outline complexity (YOLOv8 outline complexity C = 0.42), which matches the characteristics of a person.

[0211] Millimeter-wave radar: tracks the target's three-dimensional trajectory at a speed of 1.2 m / s and a trajectory curvature of 0.18 rad / m. The trajectory point coordinate sequence shows the target moving slowly along the fence.

[0212] Weather sensor data: real-time temperature 25℃, humidity 68%, wind speed 3m / s, solar radiation flux 950W / m² 2 .

[0213] 2. Dynamic threshold generation for edge computing nodes

[0214] Baseline threshold: Advanced protection level T b =1.8.

[0215] Compensation factor calculation: No rain (R=0), therefore comp1=1.0.

[0216] Thermal deposition compensation factor: Surface material is asphalt (G = ASPHALT), solar radiation flux S = 950 W / m 2 >800, temperature T=25℃, then:

[0217]

[0218] Conventional environmental compensation factor: During sunny daytime hours (14:00 falls within the 7-18h daytime period), comp3 = 1.35 is obtained from the table.

[0219] Final dynamic threshold: T d=1.8×1.0×1.175×(1+0.01×max(0,3-5))=1.8×1.175=2.115

[0220] 3. Target Classification and Behavior Analysis on Cloud Analytics Platform

[0221] Multimodal target classifier:

[0222] Infrared characteristics: Hot spots are concentrated on the torso (personnel characteristics).

[0223] Visible light characteristics: Aspect ratio 0.95, moderate outline complexity (personnel).

[0224] Radar characteristics: speed 1.2 m / s, trajectory curvature 0.18 rad / m (walking personnel).

[0225] Overall assessment: The target is "personnel".

[0226] LSTM Behavior Analyzer:

[0227] Temporal feature extraction: Dwell index SI = 0.25 (5 seconds of stillness within a total time of 20 seconds); Path complexity PC = 0.6 (detour around the fence).

[0228] Network output: p 路过 =0.15, p 观察 =0.72, p 入侵 =0.13.

[0229] Behavior determination: The target is in the "observation" phase, triggering an intermediate response.

[0230] 4. Response execution layer coordinated handling

[0231] Audible and visual alarm device: Activates high-frequency alarm (120dB) and flashes warning light to deter targets.

[0232] Unmanned Aerial Vehicle (UAV) dispatch system: dispatches patrol UAVs (equipped with infrared pods) for close-range reconnaissance and transmits high-definition images back in real time.

[0233] Access control / lighting linkage: The searchlights near the north fence turn on automatically, and the fence gate is remotely locked.

[0234] Security intervention: Ground security personnel arrive at the scene within 3 minutes of receiving the alarm, and drones guide the interception.

[0235] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A perimeter intrusion detection system based on the synergy of infrared thermal imaging and intelligent algorithms, characterized in that, It includes a multi-source perception layer, edge computing nodes, cloud analytics platform, and response execution layer; The multi-source sensing layer is equipped with an infrared thermal imager, a visible light camera, a millimeter-wave radar, and a meteorological sensor to simultaneously acquire multi-dimensional data; The edge computing node is equipped with a dynamic threshold generation module, which calculates compensation factors in real time through the multi-dimensional data to generate dynamic thresholds. The cloud analytics platform is equipped with a multimodal target classifier and an LSTM behavior analyzer. The multimodal target classifier is used to classify targets, and the LSTM behavior analyzer is used to predict the intent of target behavior. The response execution layer triggers a tiered response based on the analysis results from the cloud analytics platform.

2. The perimeter intrusion detection system based on the synergy of infrared thermal imaging and intelligent algorithms according to claim 1, characterized in that, The multi-dimensional data includes the temperature difference matrix of the target captured by the infrared thermal imager, the RGB image of the target captured by the visible light camera, the motion trajectory of the target captured by the millimeter-wave radar, and the ambient temperature, humidity, solar radiation flux, wind speed, and precipitation rate captured by the meteorological sensor.

3. The perimeter intrusion detection system based on the synergy of infrared thermal imaging and intelligent algorithms according to claim 2, characterized in that, The multimodal target classifier extracts the infrared thermal features, visible light morphological features, and radar motion features of the target through the multidimensional data. The multimodal target classifier classifies the target based on the infrared thermal features, visible light morphological features, and radar motion features, combined with the dynamic threshold. The target categories are divided into personnel, vehicles, drones, and animals.

4. The perimeter intrusion detection system based on the synergy of infrared thermal imaging and intelligent algorithms according to claim 3, characterized in that, The LSTM behavior analyzer is used to predict the intent of target behavior. The output results are divided into passing by, observing, and intruding.

5. The perimeter intrusion detection system based on the synergy of infrared thermal imaging and intelligent algorithms according to claim 4, characterized in that, The hierarchical response of the response execution layer includes audible and visual alarm prompts, drone tracking, and security dispatch.

6. The perimeter intrusion detection system based on the synergy of infrared thermal imaging and intelligent algorithms according to claim 5, characterized in that, The final dynamic threshold T of the dynamic threshold generation module d Based on the benchmark threshold T b The compensation factor comp and the wind speed correction term are jointly determined, and the formula is: T d =T b ·comp·(1+0.01·max(0,v w -5)) Where: v w The ambient wind speed is captured by the meteorological sensor. Wind speed correction term: wind speed v w When the speed is >5m / s, wind vibration will cause fluctuations in the infrared signal, and the threshold needs to be linearly amplified to suppress noise. Baseline threshold T b Based on the preset protection level, standard level T b =2.5, Advanced Level T b =1.8, Special Grade T b =1.2; The compensation factor comp is divided into rainy day compensation factor comp1, thermal accumulation compensation factor comp2 and normal environment compensation factor comp3 according to the scenario. The formula for the rain compensation factor comp1 is: comp1 = 1.2 + 0.05·R Where: R is the precipitation rate captured by the meteorological sensor, in mm / h; the rain compensation factor trigger condition is: the meteorological sensor detects a precipitation rate R>0; The formula for the thermal accumulation compensation factor comp2 is: Where: T is the ambient temperature captured by the meteorological sensor, in degrees Celsius; the triggering condition for the thermal accumulation compensation factor is: the surface material G = ASPHALT (asphalt) and the solar radiation flux S captured by the meteorological sensor > 800 W / m; The conventional environmental compensation factor comp3 is obtained by looking up historical experience data in a table. The triggering condition for the conventional environmental compensation factor is a non-rainy day and a non-thermal deposition environment.

7. The perimeter intrusion detection system based on the synergy of infrared thermal imaging and intelligent algorithms according to claim 6, characterized in that, The classification method of the multimodal target classifier includes the following steps: Step 1: Detect the thermal radiation intensity of the target surface using the infrared thermal imager and extract the following key features: Hot spot distribution histogram: The target hot spot energy is divided into 10 intervals in the spatial domain, and the energy proportion of each interval is statistically analyzed to describe the spatial distribution uniformity of the hot spot. Maximum temperature difference: Calculates the difference between the highest and lowest temperatures on the target surface (unit: °C), reflecting the degree of concentration of heat sources inside the target; Thermal radiation area percentage: The ratio of the area covered by the hot spot to the total projected area of ​​the target (unit: %), used to distinguish the relative size of the heat-generating area; The visible light camera extracts the morphological features of the target using target detection and image analysis algorithms. YOLOv8 contour complexity: Used to quantify the complexity of a target structure, calculated based on the frequency of curvature changes in the target edge contour, using the following formula: Where: L is the total length of the profile, θ i Let be the tangent angle at the i-th edge point; Aspect ratio: The ratio of the target's height to its width, reflecting the target's geometric shape; The millimeter-wave radar extracts the kinematic features of the target through the Doppler effect and target tracking algorithms: Trajectory curvature: The degree of curvature of the target motion trajectory, defined as the rate of change of the tangent direction of the path, used to distinguish between linear motion and directional motion; Speed ​​range: The instantaneous speed of the target's movement (unit: m / s), reflecting the target's motion capability; Step 2: Initial screening of radar motion characteristics, coarse classification based on dynamic constraints. Velocity < 1 m / s and trajectory curvature > 0.5 rad / m: Preliminary assessment as a candidate set for "animal"; Speed ​​≤ 1 m / s ≤ 5 m / s and trajectory curvature ≤ 0.3 rad / m: Preliminary assessment as a candidate set for "personnel / drone". Speed ​​≥ 5 m / s and trajectory curvature ≤ 0.3 rad / m: Preliminarily identified as a candidate set for "vehicle"; Step 3: Infrared thermal property re-screening, fine classification based on thermodynamic constraints. "Animal" candidate set: If the maximum temperature difference is <5℃, the "animal" category is pre-confirmed; Candidate set for "personnel / drone swarms": If the maximum temperature difference is ≥5℃ and the hot spot distribution histogram is ≈30%-50%, it is initially determined to be a candidate set for "personnel" and will proceed to the next step of visible light morphology verification; If the maximum temperature difference is ≥5℃ and the hot spot distribution histogram is ≥60%, it is initially determined to be a candidate set for "drone swarms" and will proceed to the next step of visible light morphology verification. Candidate set for "vehicles"; if the maximum temperature difference is ≥5℃ and the hot spot distribution histogram is <30%, the "vehicle" category is pre-confirmed. Step 4: Visible light morphology verification, final confirmation based on geometric features. Candidate set for "personnel": Moderate profile complexity, confirming the "personnel" category; Candidate set for "drone": Aspect ratio ≈ 0.3, confirming the "drone" category; Candidate set for "vehicle": highest outline complexity, aspect ratio > 1.0, confirming the "vehicle" category; "Animal" candidate set: lowest outline complexity, no fixed aspect ratio, confirming the "animal" category.

8. The perimeter intrusion detection system based on the synergy of infrared thermal imaging and intelligent algorithms according to claim 7, characterized in that, The LSTM behavior analyzer takes the target's motion trajectory in the spatiotemporal dimension as input data, and the specific processing flow includes the following steps: Step 1: Data Input and Preprocessing Data Acquisition: Obtain the target's three-dimensional coordinate sequence {(x1, y1, t1), (x2, y2, t2), ..., (x...} using millimeter-wave radar. n y n y n The sampling frequency is 10Hz. Data cleaning: outliers are removed, and missing values ​​in the time window are filled using linear interpolation; Normalization: Mapping coordinate values ​​to an interval, the formula is: Ensure that data of different dimensions are trained on the same scale; Window partitioning: The sequence is divided according to the time step Δt = 1 second, and a sliding window is constructed to form a supervised learning dataset; Step 2: Feature Extraction Key temporal features are extracted from the original trajectory. These key temporal features include basic features and higher-order features. The basic features include velocity, acceleration, and trajectory curvature. The higher-order features include dwell index and path complexity. The dwell index SI, representing the percentage of time the target remains stationary, is calculated using the following formula: Where: δ is an indicator function, which takes the value 1 when the condition is met and 0 otherwise; the threshold of 2 seconds is used to distinguish between a brief stop and passing by. The path complexity PC is the ratio of the actual path length to the straight-line distance, and its formula is: Where: PC value range is [0,1], the larger the value, the more circuitous the path, which may be an observation or exploratory behavior; Step 3: Network Architecture and Computation Process The LSTM behavior analyzer employs a bidirectional long short-term memory network combined with an attention mechanism to capture temporal dependencies and focus on key behavioral segments. Its specific structure is as follows: Input layer: Receives the key temporal features obtained in the second step; Bidirectional LSTM layer: Forward LSTM: Processes the sequence in chronological order and outputs the hidden state. Reverse LSTM: Processes the sequence in reverse order and outputs the hidden state. Feature fusion: concatenating the bidirectional hidden states into Capture global time-series dependencies; Attention mechanism: Calculate the attention weight α at each time step t The formula is: Among them: W q A learnable query matrix used to measure the importance of time steps; Generate context vectors Focus on key behavioral time steps; Fully connected layer: Maps the context vector c to a 3D output space, and obtains the behavior probability distribution through the Softmax function. P=Softmax(W o ·c+b o ) Among them: W o To output the weight matrix, b o As a bias term, the output P = {p 路过 p 观察 p 入侵 The sum of probabilities is 1. Step 4: Output and Decision Logic Single-frame decision: If p at a certain time step 入侵 If the value is >0.6, it is marked as "suspected intrusion" and a primary warning (audio and visual alarm) is triggered. Timing verification: If p for 3 consecutive time steps 入侵 A value >0.7 indicates "confirmed intrusion," triggering an advanced response (drone tracking, security dispatch). Intentional grading: If p 观察 A value >0.5 is considered "malicious observation". 路过 A value >0.5 is considered "harmless passerby," assisting security personnel in decision-making.

Images