Invader detection method and device applied to perimeter security system

Abstract

The invention discloses an invading object detection method and device applied to a perimeter security system, and relates to the technical field of invading object detection.According to the invading object detection method and device, a background model is dynamically updated by combining adaptive background modeling with time sequence analysis, background stability is kept in dynamic environments such as illumination variation and leaf shaking, and the invading object detection efficiency is improved. Different types of background changes are distinguished through a Gaussian mixture model, and the coexistence condition of long-term static background and short-time dynamic change can be adapted; in the target detection stage, a front-and-back frame difference method and a background subtraction method are combined, a moving target is extracted by utilizing complementary characteristics of the two methods, in the target classification and behavior analysis stage, spatial morphological characteristics of the target are extracted, the motion trend of the target is analyzed and predicted in combination with a time sequence, and track prediction is performed by adopting Kalman filtering. Even if the target is shielded for a short time or moves slowly, the tracking stability can still be kept.

Description

Technical Field

[0001] The present invention relates to the technical field of intrusion detection, and in particular to an intrusion detection method and device applied to a perimeter security system. Background Art

[0002] Most of the vehicle yards in the wild are equipped with intrusion detection devices to ensure safety. Traditional intrusion detection technologies mainly rely on video surveillance, motion detection, and target detection algorithms, or use the frame difference method and background subtraction technology to determine whether there is an intrusion into the monitored area by analyzing the changes between video frames.

[0003] However, the environment of the vehicle yard in the wild is complex, and there are a large number of dynamic interference factors in the background, such as wind-blown leaves, light and shadow changes, dust, etc. These factors may cause false alarms, making it difficult for the detection system to accurately distinguish real intrusions from environmental noise; in addition, static backgrounds such as trees and vehicles in the vehicle yard may be similar in shape to the intrusion target, further increasing the risk of misjudgment; in low-light or long-distance monitoring scenarios, the contrast between the target and the background decreases, and traditional target detection algorithms such as YOLO only focus on spatial information and are difficult to accurately extract the boundaries, resulting in a decrease in detection accuracy.

[0004] To reduce the background false detection rate, some traditional methods introduce background modeling, motion detection optimization, and multi-modal fusion strategies. The background modeling method continuously learns the scene changes and constructs a stable background model in the areas where no target has been detected for a long time to reduce false alarms, but it is still prone to failure in environments such as the rustling of grass and water wave reflections; while the motion detection optimization relies on temporal information for analysis, such as using the optical flow method or frame matching before and after to distinguish moving objects from static background interference, but when the intrusion target moves slowly or is occluded, the detection effect significantly decreases; it can be seen that in a complex background environment, there is still a high false detection rate, and there is an urgent need for an intrusion detection method and device applicable to the perimeter security system of the vehicle yard in the wild to improve the detection accuracy and reduce the influence of environmental interference. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides an intrusion detection method and device applied to a perimeter security system to solve the problems of high false detection rate of traditional intrusion detection methods in complex backgrounds, being affected by dynamic interference, low contrast, and target occlusion, insufficient detection stability, and difficulty in effectively identifying real intrusions.

[0007] To solve the above technical problems, the present invention provides the following technical solutions:

[0008] In a first aspect, an embodiment of the present invention provides an intrusion detection device applied to a perimeter security system, which includes,

[0009] An environmental perception module, which is used to collect real-time image data of a monitoring area through a visible light camera and a thermal imaging sensor, and perform noise filtering and preprocessing on the image data;

[0010] A target detection and optimization module, which, based on the updated background model, uses the front and rear frame difference method to detect the dynamic change area in the scene, and combines the background subtraction method to eliminate the long-term stable background information and extract the moving target area;

[0011] A target classification and behavior analysis module, which, based on the optimized moving target area, extracts the spatial form and motion trajectory characteristics of the intruder, and predicts the target motion trend based on time series analysis to generate classification and behavior analysis results;

[0012] During the classification and behavior analysis process, a density clustering analysis method is used to perform anomaly detection on the motion trajectory, identify the intruder category, and determine whether there is abnormal behavior according to the motion characteristics and spatio-temporal position of the intruder;

[0013] An alarm trigger and response module, which, based on the classification and behavior analysis results, adjusts the target detection strategy and triggers an alarm when the preset alarm conditions are met.

[0014] As a preferred solution of the intruder detection device applied to the perimeter security system of the present invention, wherein: in the environmental perception module, an adaptive background modeling method is used to construct a background model, dynamically learn the areas where intruders have not been detected for a long time, and update the background model based on time series analysis. The updated background model is used as the input of the target detection and optimization module.

[0015] As a preferred solution of the intruder detection device applied to the perimeter security system of the present invention, wherein: during the extraction process of the moving target area, an adaptive contrast enhancement technology is used to improve the distinguishability of the target area, and the target boundary is optimized based on edge information. The optimized moving target area is used as the input of the target classification and behavior analysis module.

[0016] In a second aspect, the present invention provides an intruder detection method applied to a perimeter security system, including,

[0017] Step S1, collect real-time image data of a monitoring area through a visible light camera and a thermal imaging sensor, and perform noise filtering and preprocessing on the image data;

[0018] Use an adaptive background modeling method to construct a background model, dynamically learn the areas where intruders have not been detected for a long time, and update the background model based on time series analysis;

[0019] Step S2: Based on the background model updated in step S1, use the frame difference method to detect the dynamic change regions in the scene, and combine the background subtraction method to eliminate the long-term stable background information and extract the moving target regions;

[0020] During the extraction process of the moving target regions, use the adaptive contrast enhancement technology to improve the recognizability of the target regions, and optimize the target boundaries based on the edge information;

[0021] Step S3: Based on the optimized moving target regions in step S2, extract the spatial form and motion trajectory features of the intruders in the moving target regions, and predict the target motion trends based on the time series analysis to generate the classification and behavior analysis results;

[0022] Step S4: According to the classification and behavior analysis results, adjust the target detection strategy and trigger an alarm when the preset alarm conditions are met.

[0023] As a preferred solution of the intruder detection method applied to the perimeter security system of the present invention, wherein: the step of using the adaptive background modeling method to construct the background model, dynamically learning the regions where no intruders have been detected for a long time, and updating the background model based on the time series analysis is as follows,

[0024] Initialize the background model, and set the input image sequence as: I t (x, y),

[0025] wherein, I t (x, y) represents the gray level or color value at the pixel point (x, y) at time t, t represents the time series index, and x, y represent the pixel coordinates in the image,

[0026] Initialize the background model B t (x, y) as the mean value of the first N frames:

[0027]

[0028] wherein, B 0 (x, y) represents the initial background model, and N is the number of initialization frames,

[0029] Use the adaptive Gaussian mixture model GMM to update the background, define the background distribution as K Gaussian components, and the probability distribution calculation formula is:

[0030]

[0031] wherein, P(I t (x, y)) represents the probability distribution of the pixel value I t (x, y), K represents the number of Gaussian distributions, w i(x, y, t) is the weight of the i-th Gaussian component, with μ i (x, y, t) as the mean, and variance of the normal distribution, where μ i (x, y, t) is the mean of the i-th Gaussian component, and

[0032] the update method for Gaussian component parameters is as follows:

[0033] If the current pixel value I t (x, y) belongs to a certain Gaussian component, then update its mean and variance. The update formulas are:

[0034] μ i (x, y, t + 1) = (1 - α)μ i (x, y, t) + αI t (x, y),

[0035]

[0036] where α is the learning rate,

[0037] If I t (x, y) does not belong to any existing Gaussian component, then replace the Gaussian component with the smallest weight:

[0038] w i (x, y, t + 1) = (1 - η)w i (x, y, t) + η,

[0039] where η is the initialization weight of the new component;

[0040] Update the background model using the time series analysis method. Adopt the exponentially weighted moving average EWMA to smooth the historical data. The update formula is:

[0041] B t (x, y) = λI t (x, y) + (1 - λ)B t-1 (x, y),

[0042] where B t (x, y) is the background model value at time t, and λ is the time series smoothing factor, with a value range of (0, 1);

[0043] For areas where no intrusion has been detected for a long time, adopt an adaptive learning rate to adjust the background update. The adjustment formula is:

[0044] λ′ = λ · exp(-γ · T stable ),

[0045] Wherein, λ′ is the learning rate after dynamic adjustment, γ is the adjustment parameter, and T stable represents the stable time when no intrusion is detected in this area.

[0046] As a preferred solution of the intrusion detection method applied to the perimeter security system described in the present invention, wherein: the steps of detecting the dynamic change area in the scene by using the front and rear frame difference method and combining the background subtraction method to eliminate the long-term stable background information and extract the moving target area are as follows,

[0047] Calculate the pixel change between the current frame and the previous frame by using the front and rear frame difference, and the calculation formula is:

[0048] D t (x,y) = |I t (x,y) - I t-1 (x,y)|,

[0049] Wherein, D t (x,y) represents the inter-frame change value of the pixel point (x,y) at time t,

[0050] Extract the moving area by using the binarization method:

[0051] If D t (x,y) > θ d , then M t (x,y) = 1, otherwise M t (x,y) = 0,

[0052] Wherein, M t (x,y) is the binary motion mask, 1 represents the moving area, 0 represents the background, and θ d is the frame difference binarization threshold;

[0053] Combine the background subtraction method to eliminate the stable background, and calculate the foreground area based on the background model B t (x,y) generated in step S1, and the calculation formula is:

[0054] F t (x,y) = |I t (x,y) - B t (x,y)|,

[0055] Wherein, F t (x,y) represents the foreground pixel intensity after background subtraction,

[0056] Extract the foreground area by using the binarization method:

[0057] If F t (x,y) > θ f , then G t (x,y) = 1, otherwise Gt (x, y) = 0,

[0058] where G t (x, y) is the foreground binary mask, and θ f is the background subtraction binarization threshold;

[0059] Combining the foreground and background frame differences and the background subtraction results, using logical operations to combine the foreground and background frame difference and background subtraction methods to remove noise points:

[0060] R t (x, y) = M t (x, y) ∩ G t (x, y),

[0061] where R t (x, y) is the finally extracted moving target area, and ∩ represents the logical AND operation;

[0062] Optimize the moving target area, and the optimization steps include:

[0063] Enhance the contrast of the extracted moving target area R t (x, y) in the following way:

[0064]

[0065] where I′ t (x, y) is the enhanced pixel value, I min , I max is the minimum and maximum pixel values of the moving area,

[0066] Calculate the edges using the Sobel operator:

[0067]

[0068] where E t (x, y) is the edge intensity, G x , G y represent the horizontal and vertical gradients respectively,

[0069] The finally optimized moving target area is O t (x, y):

[0070] O t (x, y) = R t (x, y) ∪ E t (x, y),

[0071] where O t (x, y) is the finally extracted target area, and ∪ represents the logical OR operation.

[0072] As a preferred solution of an intrusion detection method for a perimeter security system according to the present invention, in the process of classification and behavior analysis, a density clustering analysis method is used to perform anomaly detection on the motion trajectory, identify the intrusion category, and determine whether there is abnormal behavior according to the motion characteristics and spatio-temporal position of the intrusion. The abnormal behaviors include:

[0073] a) The target enters a preset sensitive area.

[0074] b) The target stays in the monitoring area for a long time exceeding the set threshold.

[0075] As a preferred solution of an intrusion detection method for a perimeter security system according to the present invention, the steps of extracting the spatial form and motion trajectory characteristics of the intrusion in the moving target area, predicting the target motion trend based on time series analysis, and generating the classification and behavior analysis results are as follows:

[0076] Calculate the geometric features of the target area, including the area A t , the centroid C t , and the aspect ratio r t . The calculation formulas are respectively:

[0077]

[0078]

[0079] Among them, A t represents the area of the target area, C t is the centroid coordinate of the target area, w t , h t is the width and height of the target area, and r t is the aspect ratio of the target area.

[0080] Use the Kalman filter to predict the target trajectory and define the state vector X t :

[0081]

[0082] The update formula is:

[0083] X t+1 = FX t + W t ,

[0084] Among them, X t is the target state, including the position (x t , y t ) and speed (v x , v y ), F is the state transition matrix, and W tis the process noise,

[0085] Use density clustering to analyze the motion trajectory, calculate the trajectory point density and classify the target:

[0086]

[0087] where Class(X t ) represents the class label of the target state X at time t t , and P i (X t ) is the trajectory probability of class i.

[0088] As a preferred solution of the intrusion detection method applied to the perimeter security system according to the present invention, wherein: in step S4, the ways to adjust the target detection strategy include:

[0089] If an intrusion is detected but the confidence level is lower than the set threshold, then adopt a multi-frame fusion strategy to enhance the target features;

[0090] If the classification and behavior analysis results meet the alarm conditions, trigger an alarm, and re-confirm the target based on the dual-modal data of visible light and thermal imaging.

[0091] As a preferred solution of the intrusion detection method applied to the perimeter security system according to the present invention, wherein: the step of adjusting the target detection strategy according to the classification and behavior analysis results and triggering an alarm when the preset alarm conditions are met is,

[0092] If an intrusion is detected but the confidence level is lower than the set threshold θ c , then adopt a multi-frame fusion strategy to enhance the target features, and the adjustment formula is:

[0093]

[0094] where I′ t (x, y) is the fused image, N is the number of fused frames,

[0095] If the target meets the alarm conditions, trigger an alarm, and based on the dual-modal data, re-confirm the target,

[0096]

[0097] where S represents the cumulative trajectory classification confidence level within the time range from t = 0 to t = T, T represents the maximum time range of trajectory analysis, and P t (X t ) represents the classification confidence level that the target state X at time t t belongs to a certain class. If S > θ s , trigger an alarm.

[0098] The beneficial effects of the present invention are as follows: The present invention utilizes adaptive background modeling combined with time series analysis to dynamically update the background model, maintains background stability in dynamic environments such as light changes and leaf shaking, and differentiates different types of background changes through the Gaussian mixture model, being able to adapt to the coexistence of long-term static backgrounds and short-term dynamic changes.

[0099] In the present invention, in the target detection stage, the frame difference method before and after is combined with the background subtraction method, and the complementary characteristics of the two are used to extract moving targets. At the same time, adaptive contrast enhancement is adopted to optimize the recognizability of the target area, and the Sobel operator is used to enhance the target edge details, enabling accurate identification of intrusions even in low-contrast environments; in the target classification and behavior analysis stage, the spatial morphological features of the target are extracted, the time series analysis is combined to predict the target movement trend, and the Kalman filter is used for trajectory prediction. Even when the target is short-term occluded or moves slowly, the tracking stability can still be maintained.

[0100] In the present invention, the density clustering method is adopted to analyze the target trajectory, differentiate different types of intrusions, and detect abnormal behaviors such as the target entering the sensitive area or staying for a long time, thereby reducing the missed detection rate of abnormal targets and improving the robustness of detection.

[0101] In the present invention, the multi-frame fusion strategy is combined to enhance the features of low-confidence targets, and the target is reconfirmed based on the dual-modal data of visible light and thermal imaging to ensure the reliability of detection and improve the real-time response ability when the alarm condition is triggered. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0103] Figure 1 It is a schematic framework diagram of the intrusion detection device of the present invention applied to the perimeter security system.

[0104] Figure 2 It is a schematic flow diagram of the intrusion detection method of the present invention applied to the perimeter security system. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0105] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific embodiments of the present invention in conjunction with the drawings in the specification.

[0106] In the following description, numerous specific details are set forth to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways different from those described herein. Persons skilled in the art can make similar extensions without departing from the spirit of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0107] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation manner of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that mutually excludes other embodiments.

[0108] Embodiment 1, referring to Figure 1 and Figure 2 , this embodiment provides an intrusion detection device applied to a perimeter security system, including:

[0109] An environmental perception module, configured to collect real-time image data of a monitoring area through a visible light camera and a thermal imaging sensor, and perform noise filtering and preprocessing on the image data;

[0110] In the environmental perception module, an adaptive background modeling method is used to construct a background model, dynamically learn areas where no intrusion has been detected for a long time, and update the background model based on time series analysis. The updated background model is used as the input to the target detection and optimization module;

[0111] A target detection and optimization module, based on the updated background model, uses the front and back frame difference method to detect dynamic change areas in the scene, and combines the background subtraction method to eliminate long-term stable background information and extract the moving target area;

[0112] During the extraction process of the moving target area, an adaptive contrast enhancement technique is used to improve the distinguishability of the target area, and the target boundary is optimized based on edge information. The optimized moving target area is used as the input to the target classification and behavior analysis module;

[0113] A target classification and behavior analysis module, based on the optimized moving target area, extracts the spatial form and motion trajectory features of the intrusion, and predicts the target motion trend based on time series analysis to generate classification and behavior analysis results;

[0114] During the classification and behavior analysis process, a density clustering analysis method is used to perform anomaly detection on the motion trajectory, identify the intrusion category, and determine whether there is abnormal behavior based on the motion characteristics and spatio-temporal position of the intrusion;

[0115] An alarm trigger and response module, based on the classification and behavior analysis results, adjusts the target detection strategy and triggers an alarm when preset alarm conditions are met.

[0116] This embodiment also provides a detection method for the intrusion object detection device applied to the perimeter security system, including:

[0117] Step S1, collect real-time image data of the monitoring area through a visible light camera and a thermal imaging sensor, and perform noise filtering and preprocessing on the image data;

[0118] Use the adaptive background modeling method to construct a background model, perform dynamic learning on the areas where intrusion objects have not been detected for a long time, and update the background model based on time series analysis;

[0119] The steps of using the adaptive background modeling method to construct a background model, performing dynamic learning on the areas where intrusion objects have not been detected for a long time, and updating the background model based on time series analysis are as follows:

[0120] Initialize the background model, and set the input image sequence as: I t (x,y),

[0121] where, I t (x,y) represents the gray or color value at the pixel point (x,y) at time t, t represents the time series index, and x,y represent the pixel coordinates in the image.

[0122] Initialize the background model B t (x,y) as the mean of the first N frames:

[0123]

[0124] where, B 0 (x,y) represents the initial background model, and N is the number of initialization frames.

[0125] Use the adaptive Gaussian mixture model GMM to update the background, define the background distribution as K Gaussian components, and the probability distribution calculation formula is:

[0126]

[0127] where, P(I t (x,y)) represents the probability distribution of the pixel value I t (x,y), K represents the number of Gaussian distributions, w i (x,y,t) is the weight of the i-th Gaussian component. is the normal distribution with μ i (x,y,t) as the mean and as the variance, μ i (x,y,t) is the mean of the i-th Gaussian component, is the variance of the i-th Gaussian component.

[0128] The updating method of Gaussian component parameters is as follows:

[0129] If the current pixel value I t (x, y) belongs to a certain Gaussian component, then its mean value and variance are updated, and the update formula is:

[0130] μ i (x, y, t + 1) = (1 - α)μ i (x, y, t) + αI t (x, y),

[0131]

[0132] where α is the learning rate,

[0133] If I t (x, y) does not belong to any existing Gaussian component, then replace the Gaussian component with the smallest weight:

[0134] w i (x, y, t + 1) = (1 - η)w i (x, y, t) + η,

[0135] where η is the initial weight of the new component;

[0136] The background model is updated using the time series analysis method, and the exponentially weighted moving average EWMA is used to smooth the historical data. The update formula is:

[0137] B t (x, y) = λI t (x, y) + (1 - λ)B t-1 (x, y),

[0138] where B t (x, y) is the background model value at time t, and λ is the time series smoothing factor, with a value range of (0, 1);

[0139] For areas where intrusions have not been detected for a long time, the background update is adjusted using an adaptive learning rate, and the adjustment formula is:

[0140] λ′ = λ · exp(-γ · T stable ),

[0141] where λ′ is the dynamically adjusted learning rate, γ is the adjustment parameter, and T stable represents the stable time when no intrusion has been detected in this area;

[0142] Specifically, here, the initial background is constructed by calculating the initial mean value, and then the Gaussian Mixture Model (GMM) is used for dynamic modeling to make the background model adaptive. GMM allows multiple Gaussian distributions to represent different types of background changes, such as illumination changes or periodically moving objects (such as swaying leaves). Subsequently, the Exponentially Weighted Moving Average (EWMA) method in time series analysis is used to smoothly update the background to make it more stable. For areas without intrusions for a long time, an exponential decay is adopted to adjust the learning rate so that the background model will not be overly updated due to short-term noise, thereby effectively coping with illumination changes and dynamic environmental changes and improving the accuracy and robustness of intrusion detection.

[0143] Step S2: Based on the background model updated in step S1, the method of differential between consecutive frames is used to detect the dynamic change regions in the scene, and the background subtraction method is combined to eliminate the long-term stable background information and extract the moving target regions.

[0144] During the extraction process of the moving target regions, the adaptive contrast enhancement technology is adopted to improve the distinguishability of the target regions, and the target boundaries are optimized based on the edge information.

[0145] The steps of using the method of differential between consecutive frames to detect the dynamic change regions in the scene, and combining the background subtraction method to eliminate the long-term stable background information and extract the moving target regions are as follows.

[0146] The pixel changes between the current frame and the previous frame are calculated using the method of differential between consecutive frames, and the calculation formula is:

[0147] D t (x,y) = |I t (x,y) - I t-1 (x,y)|,

[0148] where D t (x,y) represents the inter-frame change value of the pixel point (x,y) at time t.

[0149] The moving regions are extracted using the binarization method:

[0150] If D t (x,y) > θ d , then M t (x,y) = 1, otherwise M t (x,y) = 0.

[0151] where M t (x,y) is the binary motion mask, 1 represents the moving region, 0 represents the background, and θ d is the frame difference binarization threshold.

[0152] The stable background is eliminated by combining the background subtraction method, and based on the background model B generated in step S1 t(x, y), calculate the foreground region, and the calculation formula is:

[0153] F t (x, y) = |I t (x, y) - B t (x, y)|,

[0154] where F t (x, y) represents the foreground pixel intensity after background subtraction,

[0155] Adopt the binarization method to extract the foreground region:

[0156] If F t (x, y) > θ f , then G t (x, y) = 1, otherwise G t (x, y) = 0,

[0157] where G t (x, y) is the foreground binary mask, and θ f is the background subtraction binarization threshold;

[0158] Integrate the foreground and background frame differences and the background subtraction results, and use logical operations to combine the foreground and background frame difference and background subtraction methods to remove noise points:

[0159] R t (x, y) = M t (x, y) ∩ G t (x, y),

[0160] where R t (x, y) is the finally extracted moving target region, and ∩ represents the logical AND operation;

[0161] Optimize the moving target region, and the optimization steps include:

[0162] Enhance the contrast of the extracted moving target region R t (x, y), and the enhancement method is:

[0163]

[0164] where I′ t (x, y) is the enhanced pixel value, I min , I max the minimum and maximum pixel values of the moving region, and use the Sobel operator to calculate the edge:

[0165]

[0166] where E t (x, y) is the edge intensity, G x,G y represent the gradients in the horizontal and vertical directions respectively, and the finally optimized moving target region is O t (x,y):

[0167] O t (x,y) = R t (x,y) ∪ E t (x,y),

[0168] where O t (x,y) is the finally extracted target region, and ∪ represents the logical OR operation;

[0169] Specifically, the frame difference method is used to detect the dynamic change region in the scene, and the background subtraction method is combined to eliminate the long-term stable background information, and finally the moving target region is extracted;

[0170] Calculate the frame difference between the previous and next frames to generate a motion region mask, and then use the background modeling method to calculate the background subtraction foreground and perform binary processing, and the two are combined to obtain the final moving target region;

[0171] Here, to improve the quality of the target region, the adaptive contrast enhancement method is used to improve the distinguishability, and the Sobel operator is used to optimize the edge information to generate the final moving target region and effectively suppress noise;

[0172] Step S3, based on the optimized moving target region in step S2, extract the spatial form and motion trajectory characteristics of the intruders in the moving target region, and predict the target motion trend based on time series analysis to generate the classification and behavior analysis results;

[0173] In the process of classification and behavior analysis, the density clustering analysis method is used to detect anomalies in the motion trajectory, identify the intruder category, and judge whether there is abnormal behavior according to the motion characteristics and spatio-temporal position of the intruder. The abnormal behaviors include:

[0174] a) The target enters the preset sensitive area,

[0175] b) The target stays in the monitoring area for a long time exceeding the set threshold,

[0176] The steps of extracting the spatial form and motion trajectory characteristics of the intruders in the moving target region and predicting the target motion trend based on time series analysis to generate the classification and behavior analysis results are as follows:

[0177] Calculate the geometric features of the target region, including the area A t 、the centroid C t 、the aspect ratio r t , and the calculation formulas are respectively:

[0178]

[0179] Among them, A t represents the area of the target region, and C t is the centroid coordinate of the target region, w t , h t are the width and height of the target region, and x t is the aspect ratio of the length and width of the target region.

[0180] The Kalman filter is used to predict the target trajectory, and the state vector X t is defined as:

[0181]

[0182] The update formula is:

[0183] X t+1 = FX t + W t ,

[0184] Among them, X t is the target state, including the position (x t , y t ) and the speed (v x , v y ), F is the state transition matrix, and W t is the process noise.

[0185] Density clustering is used to analyze the motion trajectory, calculate the trajectory point density, and classify the target:

[0186]

[0187] Among them, Class(X t ) represents the class identifier of the target state X t at time t, and P i (X t ) is the trajectory probability of class i;

[0188] Specifically, here, spatial morphological features including area, centroid, and aspect ratio are extracted based on the optimized moving target region, and the Kalman filter method is used to calculate the target motion trajectory;

[0189] In addition, the density clustering method is used to analyze the target trajectory, classify the intruders, effectively identify different types of intruders, and combine time series analysis to predict the future motion trend of the target, improving the accuracy of classification and behavior analysis;

[0190] Step S4, according to the classification and behavior analysis results, adjust the target detection strategy, and trigger an alarm when the preset alarm condition is met;

[0191] In step S4, the ways to adjust the target detection strategy include:

[0192] If an intruder is detected but the confidence level is lower than the set threshold, then a multi-frame fusion strategy is adopted to enhance the target features;

[0193] If the classification and behavior analysis results meet the alarm conditions, an alarm is triggered, and target reconfirmation is performed on the detection results based on the visible light and thermal imaging dual-modal data;

[0195] The steps of adjusting the target detection strategy according to the classification and behavior analysis results and triggering an alarm when the preset alarm conditions are met are as follows

[0196] If an intruder is detected but the confidence level is lower than the set threshold θ c , then a multi-frame fusion strategy is adopted to enhance the target features, and the adjustment formula is:

[0197]

[0198] where I′ t (x, y) is the fused image, N is the number of fused frames,

[0199] If the target meets the alarm conditions, an alarm is triggered, and target reconfirmation is performed based on the dual-modal data

[0200]

[0201] where S represents the cumulative trajectory classification confidence level within the time range from t = 0 to t = T, T represents the maximum time range of trajectory analysis, and P t (X t ) represents the classification confidence level that the target state X t belongs to a certain category at time t. If > θ s , an alarm is triggered;

[0202] Specifically, the detection strategy is dynamically adjusted according to the target classification results, multi-frame fusion is used to enhance low-confidence targets, and if the target meets the preset conditions, reconfirmation is performed based on the visible light and thermal imaging data to improve the accuracy of alarm triggering.

[0203] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. An intrusion detection device used in a perimeter security system, characterized in that: include, An environmental perception module is used to collect real-time image data of the monitored area through a visible light camera and a thermal imaging sensor, and to perform noise filtering and preprocessing on the image data; The target detection and optimization module uses the previous and next frame difference method to detect the dynamically changing areas in the scene based on the updated background model, and combines the background subtraction method to eliminate the long-term stable background information and extract the moving target area; The target classification and behavior analysis module extracts the spatial form and motion trajectory characteristics of the intruder based on the optimized moving target area, and predicts the target movement trend based on time series analysis to generate classification and behavior analysis results; In the classification and behavior analysis process, a density cluster analysis method is used to detect abnormalities in the motion trajectory, identify the category of the intruder, and determine whether there is abnormal behavior based on the motion characteristics and spatiotemporal position of the intruder; The alarm triggering and response module adjusts the target detection strategy based on the classification and behavior analysis results, and triggers an alarm when the preset alarm conditions are met.

2. An intrusion detection device for use in a perimeter security system as claimed in claim 1, characterized in that: In the environmental perception module, an adaptive background modeling method is used to construct a background model, dynamically learn areas where no intruders are detected for a long time, and update the background model based on time series analysis. The updated background model serves as the input of the target detection and optimization module.

3. The intrusion detection device for use in a perimeter security system according to claim 2, characterized in that: In the process of extracting the moving target area, an adaptive contrast enhancement technique is used to improve the recognizability of the target area, and the target boundary is optimized based on edge information. The optimized moving target area is used as the input of the target classification and behavior analysis module.

4. An intrusion detection method applied to a perimeter security system, based on an intrusion detection device applied to a perimeter security system according to any one of claims 1 to 3, characterized in that: include: Step S1, collecting real-time image data of the monitored area through a visible light camera and a thermal imaging sensor, and performing noise filtering and preprocessing on the image data; The background model is constructed using the adaptive background modeling method, the areas where no intruders are detected for a long time are dynamically learned, and the background model is updated based on time series analysis; Step S2, based on the background model updated in step S1, the front and back frame difference method is used to detect the dynamically changing area in the scene, and the long-term stable background information is eliminated in combination with the background subtraction method to extract the moving target area; In the process of extracting the moving target area, an adaptive contrast enhancement technique is used to improve the recognizability of the target area, and the target boundary is optimized based on edge information; Step S3, based on the optimized moving target area in step S2, extract the spatial form and movement trajectory characteristics of the intruder in the moving target area, and predict the target movement trend based on time series analysis to generate classification and behavior analysis results; Step S4, adjust the target detection strategy according to the classification and behavior analysis results, and trigger an alarm when the preset alarm conditions are met.

5. The intrusion detection method for a perimeter security system as claimed in claim 4, characterized in that: The steps of constructing a background model using an adaptive background modeling method, dynamically learning an area where no intruders have been detected for a long time, and updating the background model based on time series analysis are as follows: Initialize the background model, and set the input image sequence to be: I t (x, y), Among them, I t (x, y) represents the grayscale or color value of the pixel point (x, y) at time t, t represents the time series index, x, v represents the pixel coordinates in the image, The background model B t (x, y) is initialized to the mean of the previous N frames: Among them, B0(x, y) represents the initial background model, N is the number of initialization frames, Adopt adaptive Gaussian mixture model GMM to update the background, define the background distribution as K Gaussian components, and the probability distribution calculation formula is: Among them, P(I t (x, y)) represents the pixel value I t The probability distribution of (x, y), K represents the number of Gaussian distributions, w i (x, y, t) is the weight of the i-th Gaussian component, μ i (x, y, t) is the mean, is the normal distribution with variance, μ i (x, y, t) is the mean of the i-th Gaussian component, is the variance of the i-th Gaussian component, The Gaussian component parameters are updated as follows: If the current pixel value I t If (x, y) belongs to a Gaussian component, then update its mean and variance. The update formula is: m i (x, y, t+1)=(1-a)μ i (x, y, t)+αI t (x,y), Among them, α is the learning rate, If I t If (x, y) does not belong to any existing Gaussian component, then replace it with the Gaussian component with the smallest weight: w i (x,y,t+1)=(1-η)w i (x,y,t)+η, Among them, η is the initialization weight of the new component; The background model is updated using the time series analysis method, and the exponentially weighted moving average EWMA is used to smooth the historical data. The update formula is: B t (x,y)=λI t (x,y)+(1-λ)B t-1 (x,y), Among them, B t (x, y) is the background model value at time t, λ is the time series smoothing factor, and its value range is (0, 1); For areas where no intruders are detected for a long time, an adaptive learning rate is used to adjust the background update. The adjustment formula is: λ′=λ·exp(-γ·T stable ), Among them, λ′ is the dynamically adjusted learning rate, Y is the adjustment parameter, T stable Indicates the stable time during which no intruders are detected in the area.

6. The intrusion detection method for a perimeter security system as claimed in claim 5, characterized in that: The steps of using the front and back frame difference method to detect the dynamically changing area in the scene, and combining the background subtraction method to remove the long-term stable background information and extract the moving target area are as follows: The pixel change between the current frame and the previous frame is calculated using the difference between the previous and next frames. The calculation formula is: D t (x,y)=|I t (x,y)-I t-1 (x,y)|, Among them, D t (x, y) represents the inter-frame change value of the pixel point (x, y) at time t. Use the binarization method to extract the motion area: If D t (x,y)>θ d , then M t (x, y) = 1, otherwise M t (x, y) = 0, Among them, M t (x, y) is a binary motion mask, 1 represents the motion area, 0 represents the background, θ d is the frame difference binarization threshold; Combined with the background subtraction method to remove the stable background, based on the background model B generated in step S1 t (x, y), calculate the foreground area, the calculation formula is: F t (x,y)=|I t (x,y)-B t (x,y)|, Among them, F t (x, y) represents the foreground pixel intensity after background subtraction, Use the binarization method to extract the foreground area: If F t (x,y)>θ f , then G t (x, y) = 1, otherwise G t (x, y) = 0, Among them, G t (x, y) is the foreground binary mask, θ f Binarization threshold for background subtraction; Combining the difference between the previous and next frames and the background subtraction results, the noise points can be removed by using logical operations combined with the difference between the previous and next frames and the background subtraction method: R t (x,y)=M t (x,y)∩G t (x,y), Among them, R t (x, y) is the final extracted moving target area, and n represents the logical AND operation; Optimize the moving target area. The optimization steps include: The extracted moving target area R t (x, y) is contrast enhanced by: Among them, I′ t (x, y) is the enhanced pixel value, I min , I max The minimum and maximum pixel values ​​of the motion area, Use Sobel operator to calculate the edge: Among them, E t (x, y) is the edge strength, G x , G y Represent the gradients in the horizontal and vertical directions respectively. The final optimized moving target area is O t (x, y): EITHER t (x, y)=R t (x, y)∪E t (x, y) Among them, O t (x, y) is the target area finally extracted, and ∪ represents a logical OR operation.

7. The intrusion detection method for a perimeter security system as claimed in claim 6, characterized in that: In the process of classification and behavior analysis, a density cluster analysis method is used to detect abnormalities in the motion trajectory, identify the category of the intruder, and determine whether there is abnormal behavior based on the motion characteristics and spatiotemporal position of the intruder. The abnormal behavior includes: a) The target enters the preset sensitive area, b) The target stays in the monitoring area for a long time exceeding the set threshold.

8. The intrusion detection method for a perimeter security system as claimed in claim 7, characterized in that: The steps of extracting the spatial form and motion trajectory characteristics of the intruder in the moving target area, predicting the target motion trend based on time series analysis, and generating classification and behavior analysis results are as follows: Calculate the geometric characteristics of the target area, including the area A t , Center of Gravity C t , aspect ratio r t , the calculation formulas are: Among them, A t represents the area of ​​the target area, C t is the centroid coordinate of the target area, w t ,h t is the width and height of the target area, r t is the aspect ratio of the target area, Use Kalman filtering to predict the target trajectory and define the state vector X t : The update formula is: X t+1 =FX t +W t , Among them, X t is the target state, including the position (x t ,y t ) and speed (v x , v y ), F is the state transfer matrix, W t is the process noise, Use density clustering to analyze motion trajectories, calculate trajectory point density and classify targets: Among them, Class(X t ) represents the target state X at time t t The category identifier, P i (X t ) is the trajectory probability of category i.

9. The intrusion detection method for a perimeter security system as claimed in claim 8, characterized in that: In step S4, the method of adjusting the target detection strategy includes: If an intruder is detected but the confidence is lower than the set threshold, a multi-frame fusion strategy is used to enhance the target features; If the classification and behavior analysis results meet the alarm conditions, an alarm is triggered and the detection results are reconfirmed based on the visible light and thermal imaging dual-modal data.

10. The intrusion detection method applied to a perimeter security system according to claim 9, characterized in that: The steps of adjusting the target detection strategy according to the classification and behavior analysis results and triggering an alarm when the preset alarm conditions are met are: If an intruder is detected but the confidence level is lower than the set threshold θ c , then a multi-frame fusion strategy is used to enhance the target features, and the adjustment formula is: Among them, I′ t (x, y) is the fused image, N is the number of fused frames, If the target meets the alarm conditions, the alarm is triggered and the target is reconfirmed based on the dual-modal data. Where S represents the cumulative trajectory classification confidence from t = 0 to t = T, T represents the maximum time range of trajectory analysis, and P t (X t ) represents the target state X at time t t The confidence level of a classification belonging to a certain category, if S>θ s , triggering an alarm.

Images