An intrusion prevention monitoring and early warning method and system

Abstract

This invention discloses an intrusion prevention monitoring and early warning method and system, belonging to the field of image recognition technology. The invention performs Gaussian filtering with a noise threshold on motion difference images to obtain a structure separation map; then, contour extraction and filtering are performed on the structure separation map to obtain the direction enhancement weights for each pixel. The pixel values ​​are multiplied by the corresponding direction enhancement weights to construct a direction enhancement map; subsequently, the vertical continuity intensity of the direction enhancement map is multiplied by the width stability coefficient to obtain a human morphology continuous response map; simultaneously, a multi-scale morphological top-hat transform is performed on the structure separation map, and the significance coefficients are obtained through cross-scale saliency point statistics to modulate the structure separation map, constructing a salient structure response map; finally, a dual-channel neural network is used to process the salient structure response map and the human morphology continuous response map to obtain the human morphology recognition result. When the recognition result is a human body, an intrusion warning is triggered. This invention effectively improves the accuracy of intrusion prevention monitoring and early warning.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically to an anti-intrusion monitoring and early warning method and system. Background Technology

[0002] In the field of intelligent security, intrusion monitoring and early warning technology is widely used in key areas such as industrial parks, factories, and borders. Its core requirement is to quickly and accurately identify intruding human targets within the monitoring range and issue timely warnings to mitigate security risks. Currently, one commonly used intrusion monitoring and early warning technology is a method based on median filtering denoising combined with traditional machine learning classifiers. This method first uses a median filtering algorithm to suppress noise in the acquired motion difference image, removing salt-and-pepper noise, random interference, and other signals to obtain a preliminarily denoised image. Then, the denoised image is converted to grayscale to extract simple grayscale and edge features. Finally, the extracted features are input into a traditional machine learning classifier, and the classification rules trained on the samples determine whether a human intrusion target exists within the monitoring range, thereby triggering the corresponding early warning operation.

[0003] However, this existing technology has significant technical problems and is difficult to meet the needs of accurate early warning in complex scenarios. Due to the denoising characteristics of median filtering, while suppressing noise, it easily destroys the integrity of the edge contour of human targets, resulting in missing grayscale and edge features in the extracted data, which cannot accurately reflect the morphological characteristics of the human body. At the same time, the feature recognition capabilities of traditional machine learning classifiers are limited, and they can only process simple grayscale and edge features. They have low distinguishability between interference targets (such as swaying trees or piles of debris) and human targets in complex monitoring environments, which easily leads to misjudgment and missed judgment, resulting in low accuracy of intrusion monitoring and early warning. Summary of the Invention

[0004] In view of the above-mentioned shortcomings in the prior art, the present invention provides an intrusion prevention monitoring and early warning method and system that solves the problem of low accuracy in intrusion prevention monitoring and early warning in the prior art.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: an intrusion prevention monitoring and early warning method, comprising:

[0006] The motion difference image is subjected to Gaussian filtering with a noise threshold to obtain a structure separation map.

[0007] Contour extraction and filtering are performed on the structure separation map to obtain the orientation enhancement weight of each pixel;

[0008] The pixel value of each pixel in the structure separation map is multiplied by the orientation enhancement weight of the corresponding pixel to obtain the pixel enhancement value, and the orientation enhancement map is constructed.

[0009] The longitudinal continuity intensity of the directional enhancement map is multiplied by the width stability coefficient to construct a continuous response map of human body shape.

[0010] Multi-scale morphological top-hat transformation is performed on the structure separation map. Significance coefficients are obtained through cross-scale salient point statistics. The structure separation map is then modulated to construct a salient structure response map.

[0011] A dual-channel neural network is used to process the salient structure response map and the continuous response map of human body shape to obtain the human body shape recognition result. When the human body shape recognition result is a human body, an intrusion warning is issued.

[0012] Furthermore, the Gaussian filtering process includes:

[0013] A Gaussian filter is applied to the motion difference image to obtain a smoothed image;

[0014] The pixel difference image is obtained by subtracting the smoothed image from the motion difference image pixel by pixel and taking the absolute value.

[0015] In a pixel difference image, when the pixel value is less than the noise threshold, the corresponding pixel is considered an interference point.

[0016] In a pixel difference image, when the pixel value is greater than or equal to the noise threshold, the corresponding pixel is a structure point;

[0017] For interference points, the pixel values ​​of the smoothed image are used as the output; for structural points, the pixel values ​​of the motion difference image are used as the output, thus obtaining the structural separation map.

[0018] Furthermore, the process of obtaining the orientation enhancement weights for each pixel includes:

[0019] Canny edge detection is performed on the structure separation map to obtain multiple edge contours;

[0020] Calculate the length of each edge contour, filter out edge contours whose length is greater than or equal to the minimum human body contour length threshold, and form a candidate contour set.

[0021] For each edge contour in the candidate contour set, calculate the average gradient direction of all pixels inside it to obtain the average gradient direction of the corresponding edge contour.

[0022] The orientation enhancement weight of each pixel is obtained based on the difference between the gradient direction of each pixel in each edge contour in the candidate contour set and the average gradient direction of the corresponding edge contour.

[0023] Furthermore, the process of obtaining the orientation enhancement weights for each pixel includes:

[0024] The gradient direction of each pixel and its neighboring pixels in each edge contour of the candidate contour set is subtracted from the average gradient direction, and the direction deviation is obtained by using an exponential decay term.

[0025] Convert the gradient directions of a pixel and its neighboring pixels into complex numbers.

[0026] The gradient direction in complex form is multiplied by the deviation from the direction, and the average value is taken to obtain the average complex vector;

[0027] The magnitude of the average complex vector is used as the orientation enhancement weight for the corresponding pixel, and the orientation enhancement weights for other pixels that do not belong to the candidate contour set are reset to 0.

[0028] Furthermore, the process of constructing a continuous response map of human body shape includes:

[0029] Using each pixel in the directional enhancement map as the center, the pixel enhancement values ​​of all pixels within the vertical window are added together to obtain the vertical continuous intensity of the corresponding pixel.

[0030] Centered on each pixel in the orientation enhancement map, within a horizontal window, count the number of pixels whose pixel enhancement value is greater than the average pixel enhancement value, and use this number of pixels as the effective width of the corresponding pixel.

[0031] The width stability coefficient of each pixel is obtained by comparing the effective width of each pixel with the pixels in the adjacent row.

[0032] Multiply the vertical continuous intensity of the same pixel by the width stability coefficient to obtain the morphological value of the corresponding pixel, and construct a continuous response map of human morphology.

[0033] Furthermore, the process of obtaining the salient structural response map includes:

[0034] Using three scale structuring elements, morphological top-hat transformations were performed on the structure separation map to obtain three scale top-hat response maps;

[0035] The global pixel mean is obtained by averaging the pixel values ​​of all pixels in the scale top cap response map.

[0036] In each scale top-hat response map, when the pixel value of a pixel is greater than the corresponding global pixel mean, the corresponding pixel is marked as a salient point;

[0037] In the three scale top-hat response maps, count the number of times the same pixel location is marked as a salient point;

[0038] Normalize the frequency of salient points to obtain the salient coefficient of each pixel;

[0039] At the same pixel point, the saliency coefficient is multiplied by the pixel value in the structure separation map to obtain the saliency structure response map.

[0040] Furthermore, the processing steps of a dual-channel neural network include:

[0041] The salient structure response map is processed using the first multi-convolutional channel to obtain the salient structure feature map;

[0042] The human body morphology continuous response map is processed using a second multi-convolutional channel to obtain the human body morphology feature map;

[0043] A cross-cooperative weighted fusion unit is used to fuse human morphological feature maps and salient structural feature maps to obtain a cooperative fusion feature map;

[0044] A deep feature extraction unit is used to extract deep features from the collaborative fusion feature map to obtain the collaborative fusion deep feature map;

[0045] Multi-layer spatial attention is applied to the collaboratively fused deep feature map using multi-layer spatial attention units to obtain a spatially enhanced feature map;

[0046] The output unit outputs human morphology recognition results based on spatially enhanced feature maps.

[0047] Furthermore, the cross-cooperative weighted fusion unit includes: a first-channel attention module, a second-channel attention module, multiplier M1, multiplier M2, and adder A1;

[0048] The first channel attention module is used to generate first channel attention based on the salient structural feature map;

[0049] The second-channel attention module is used to generate second-channel attention based on the human morphological feature map;

[0050] Multiplier M1 is used to multiply the first channel attention with the human body shape feature map element-wise to obtain the human body shape channel enhanced feature map;

[0051] Multiplier M2 is used to element-wise multiply the second channel attention with the salient structure feature map to obtain the salient structure channel enhanced feature map;

[0052] Adder A1 is used to add the enhanced feature map of the human body shape channel and the enhanced feature map of the salient structure channel element by element to obtain a synergistic fusion feature map.

[0053] Furthermore, the multi-layer spatial attention unit includes: a first spatial attention module, a second spatial attention module, a multiplier M3, a multiplier M4, and an adder A2;

[0054] The first spatial attention module is used to generate first spatial attention on the collaboratively fused deep feature map;

[0055] The second spatial attention module is used to generate second spatial attention from the collaboratively fused deep feature map;

[0056] Multiplier M3 is used to element-wise multiply the first spatial attention with the collaborative fusion deep feature map to obtain the first spatial augmentation map;

[0057] Multiplier M4 is used to element-wise multiply the second spatial attention with the collaborative fusion deep feature map to obtain the second spatial augmentation map;

[0058] Adder A2 is used to add the first spatial augmentation map and the second spatial augmentation map element by element to obtain a spatial augmentation feature map.

[0059] An intrusion prevention monitoring and early warning system includes: a filtering subsystem, a directional enhancement weight acquisition subsystem, a first construction subsystem, a second construction subsystem, a third construction subsystem, and an identification and early warning subsystem;

[0060] The filtering subsystem is used to perform Gaussian filtering with a noise threshold on the motion difference image to obtain a structure separation map;

[0061] The orientation enhancement weight acquisition subsystem is used to extract and filter the contour of the structure separation map and obtain the orientation enhancement weight of each pixel.

[0062] The first construction subsystem is used to multiply the pixel value of each pixel in the structure separation map with the orientation enhancement weight of the corresponding pixel to obtain the pixel enhancement value and construct the orientation enhancement map;

[0063] The second construction subsystem is used to multiply the longitudinal continuity intensity of the orientation enhancement map with the width stability coefficient to construct a continuous response map of human body shape.

[0064] The third construction subsystem is used to perform multi-scale morphological top-hat transformation on the structure separation map. By statistically analyzing cross-scale salient points, it obtains significance coefficients, modulates the structure separation map, and constructs a salient structure response map.

[0065] The identification and early warning subsystem uses a dual-channel neural network to process the salient structure response map and the continuous response map of human body shape to obtain the human body shape recognition result. When the human body shape recognition result is a human body, an intrusion warning is issued.

[0066] The beneficial effects of this invention are as follows:

[0067] 1. This invention uses Gaussian filtering with a noise threshold, which effectively suppresses salt-and-pepper noise, random interference, etc., while fully preserving the edge contours and detailed textures of the human target, avoiding the destruction of human morphological features by traditional median filtering.

[0068] 2. This invention extracts and filters contours, obtains the direction enhancement weight of each pixel, modulates the structure separation map, constructs a direction enhancement map, strengthens the response consistent with the human contour direction, suppresses the messy direction response of irrelevant interference, makes the human target more prominent in the feature map, and improves the distinction between the target and the background.

[0069] 3. This invention constructs a continuous response map of human body shape by fusing longitudinal continuity intensity and lateral width stability coefficient. Longitudinal accumulation enhances the continuity of the human body in the vertical direction, while the lateral stability coefficient constrains the smoothness of changes in human body width. This makes the response map both continuous and consistent with the morphological characteristics of the real human body, effectively filtering out interfering targets with abnormal shapes and providing clear spatial morphological features for the neural network.

[0070] 4. This invention constructs a salient structural response map through multi-scale morphological top-hat transformation and cross-scale salient point statistics. This map focuses on regions with high local contrast and prominent structures in the image, and can capture the key structural features of human targets. It complements the continuous response map of human morphology, characterizing the target from two dimensions, and further enhancing the richness and recognizability of features.

[0071] 5. This invention employs a dual-channel neural network to process the salient structure response map and the human body morphology continuous response map separately. Through dual-path feature fusion, it simultaneously learns the features of two dimensions of the target, which greatly improves the ability to distinguish between interference targets and human targets in complex scenes, effectively reduces the false positive and false negative rates, and significantly improves the accuracy and reliability of intrusion prevention monitoring and early warning. Attached Figure Description

[0072] Figure 1 A flowchart of an intrusion prevention monitoring and early warning method;

[0073] Figure 2 This is a flowchart of the processing of a dual-channel neural network;

[0074] Figure 3 This is a schematic diagram of the cross-cooperative weighted fusion unit.

[0075] Figure 4 This is a schematic diagram of the structure of a multi-layer spatial attention unit;

[0076] Figure 5 This is a schematic diagram of the structure of the first and second multi-convolutional channels;

[0077] Figure 6 This is a schematic diagram of the deep feature extraction unit. Detailed Implementation

[0078] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0079] Example 1, as Figure 1 As shown, an intrusion prevention monitoring and early warning method includes:

[0080] The motion difference image is subjected to Gaussian filtering with a noise threshold to obtain a structure separation map.

[0081] Contour extraction and filtering are performed on the structure separation map to obtain the orientation enhancement weight of each pixel;

[0082] The pixel value of each pixel in the structure separation map is multiplied by the orientation enhancement weight of the corresponding pixel to obtain the pixel enhancement value, and the orientation enhancement map is constructed.

[0083] The longitudinal continuity intensity of the directional enhancement map is multiplied by the width stability coefficient to construct a continuous response map of human body shape.

[0084] Multi-scale morphological top-hat transformation is performed on the structure separation map. Significance coefficients are obtained through cross-scale salient point statistics. The structure separation map is then modulated to construct a salient structure response map.

[0085] A dual-channel neural network is used to process the salient structure response map and the continuous response map of human body shape to obtain the human body shape recognition result. When the human body shape recognition result is a human body, an intrusion warning is issued.

[0086] In this embodiment, two types of monitoring images are first acquired: the monitoring image at the current moment and the background monitoring image when the screen is static (i.e., the monitoring background image); by calculating the pixel value difference between the monitoring image at the current moment and the static background monitoring image, and taking the absolute value of the difference, the motion difference image to be processed is obtained.

[0087] In this embodiment, the Gaussian filtering process includes:

[0088] A Gaussian filter is applied to the motion difference image to obtain a smoothed image;

[0089] The pixel difference image is obtained by subtracting the smoothed image from the motion difference image pixel by pixel and taking the absolute value.

[0090] In a pixel difference image, when the pixel value is less than the noise threshold, the corresponding pixel is considered an interference point.

[0091] In a pixel difference image, when the pixel value is greater than or equal to the noise threshold, the corresponding pixel is a structure point;

[0092] For interference points, the pixel values ​​of the smoothed image are used as the output; for structural points, the pixel values ​​of the motion difference image are used as the output, thus obtaining the structural separation map.

[0093] In this embodiment, ,in, This is a structural separation diagram with noise constraints. For Gaussian kernel, For convolution operations, For indicator functions, For motion difference images, This is the noise threshold. That is, when a pixel is a structured point (difference ≥ threshold), the original value is used; when a pixel is a noise point (difference < threshold), a smoothed value is used. Indicator function. The value is 1 if the condition is true, and 0 if the condition is false.

[0094] noise threshold It can be determined through experiments or experience.

[0095] The Gaussian filtering process of this invention effectively separates structural information from noise components in an image by introducing a noise threshold. During processing, structural points retain their original pixel values, while interference points use pixel values ​​from a smoothed image, thereby reducing the impact of noise and preserving important features in the image.

[0096] In this embodiment, the process of obtaining the orientation enhancement weights for each pixel includes:

[0097] Canny edge detection is performed on the structure separation map to obtain multiple edge contours;

[0098] Calculate the length of each edge contour, filter out edge contours whose length is greater than or equal to the minimum human body contour length threshold, and form a candidate contour set.

[0099] For each edge contour in the candidate contour set, calculate the average gradient direction of all pixels inside it, that is, the average gradient direction of all pixels on each edge contour, to obtain the average gradient direction of the corresponding edge contour.

[0100] The orientation enhancement weight of each pixel is obtained based on the difference between the gradient direction of each pixel in each edge contour in the candidate contour set and the average gradient direction of the corresponding edge contour.

[0101] In this embodiment, the length of each edge contour can be obtained by statistically summing the Euclidean distances between adjacent pixels.

[0102] In this embodiment, the process of obtaining the orientation enhancement weights for each pixel includes:

[0103] The gradient direction of each pixel and its neighboring pixels in each edge contour of the candidate contour set is subtracted from the average gradient direction, and the direction deviation is obtained by using an exponential decay term.

[0104] Convert the gradient directions of a pixel and its neighboring pixels into complex numbers.

[0105] The gradient direction in complex form is multiplied by the deviation from the direction, and the average value is taken to obtain the average complex vector;

[0106] The magnitude of the average complex vector is used as the orientation enhancement weight for the corresponding pixel, and the orientation enhancement weights for other pixels that do not belong to the candidate contour set are reset to 0.

[0107] For the selected candidate contours, the deviation of each pixel and its neighboring pixels from the average gradient direction of the corresponding contour is combined, and the directional deviation is obtained through an exponential decay term. The gradient direction is converted into a complex number and multiplied with the directional deviation, and the average value is taken. The magnitude of the resulting average complex vector is used as the directional enhancement weight of the pixel. The directional enhancement weight of pixels in non-candidate contours is uniformly set to 0.

[0108] In this embodiment, the formula for calculating the directional deviation is: ,in, For the candidate contour set, the first The first edge contour Within the neighborhood of the nth pixel Directional deviation of each pixel For the candidate contour set, the first The first edge contour Within the neighborhood of the nth pixel (including the pixel itself), the nth pixel... Gradient direction of each pixel For the candidate contour set, the first The average gradient direction of the edge contour, This is the adjustment parameter for the angle deviation. It is an exponential function. It is an absolute value.

[0109] In this embodiment, the formula for calculating the directional enhancement weight is: ,in, For the candidate contour set, the first The first edge contour directional weight enhancement for each pixel For modulo operation, For imaginary units, For the candidate contour set, the first The first edge contour Within the neighborhood of the nth pixel (including the pixel itself), the nth pixel... Gradient direction of each pixel The number of pixels in the neighborhood. For numbering.

[0110] In this embodiment, the adjustment parameter of the angle deviation Set a fixed angle range, for example, 10° to 30° (or the corresponding radian value).

[0111] In this embodiment, the minimum human body contour length threshold can be preset according to the actual needs of the monitoring scenario, or determined by statistical analysis of the human body target contour length in historical monitoring data.

[0112] This invention filters edge contours in a structural separation map, first establishing candidate contours with a certain degree of continuity. The average gradient direction of each contour is used as the overall directional benchmark. The gradient directions of pixels and their neighborhoods are compared with this benchmark direction. An exponential decay function modulates the degree of directional deviation, giving higher weights to pixels with more consistent directions. Simultaneously, the neighborhood gradient directions are converted to complex numbers and weighted averaged. The magnitude of the complex vector superposition result represents local directional consistency, thereby reducing the influence of single-point noise and directional fluctuations. This ensures that the resulting directional enhancement weights reflect the continuous directional characteristics of the contour structure. The directional enhancement weights range from 0 to 1; a larger value indicates higher directional consistency in the pixel's neighborhood.

[0113] In this embodiment, the process of constructing a continuous response map of human body shape includes:

[0114] Using each pixel in the directional enhancement map as the center, the pixel enhancement values ​​of all pixels within the vertical window are added together to obtain the vertical continuous intensity of the corresponding pixel.

[0115] Centered on each pixel in the orientation enhancement map, within a horizontal window, count the number of pixels whose pixel enhancement value is greater than the average pixel enhancement value (the average pixel enhancement value of all pixels in the orientation enhancement map), and use this number of pixels as the effective width of the corresponding pixel.

[0116] The width stability coefficient of each pixel is obtained by normalizing and mapping the difference between each pixel and the pixels in the adjacent row in the effective width.

[0117] Multiply the vertical continuous intensity of the same pixel by the width stability coefficient to obtain the morphological value of the corresponding pixel, and construct a continuous response map of human morphology.

[0118] In this embodiment, the expression for longitudinal continuous strength is:

[0119] ,

[0120] in, For the first Vertical continuous intensity of each pixel For the first The x-coordinate of each pixel For the first The ordinate of each pixel For the dimensions of the vertical window, For vertical indexing, For the first In the vertical window of the nth pixel, the first The pixel enhancement value for each pixel.

[0121] by Centered on the horizontal axis, with a fixed horizontal coordinate, the vertical window is as follows: h is used to depict the degree of continuous extension of the human body in the vertical direction. It should be slightly less than half of the minimum height of the human body so that the window can cover the local continuous structure of the human body but does not excessively cross different targets. For example, when the minimum height of the human body is about 80 pixels, h≈4~12 can be taken; when the minimum height of the human body is about 120 pixels, h≈6~18 can be taken.

[0122] by Centered on the vertical axis, with a fixed vertical axis, the horizontal window is: , The dimensions of the horizontal window. To reflect the stability of the horizontal width of the human body, the value should be close to half of the minimum width of the human body, so that the window can cover the width of the shoulders or torso while avoiding crossing adjacent targets. For example, when the minimum width of the human body is about 30 pixels, w≈9~18 can be used; when the minimum width of the human body is about 40 pixels, w≈12~24 can be used.

[0123] The formula for calculating the width stability coefficient is:

[0124] ,

[0125] in, For the first Width stability coefficient per pixel for Effective width, for Effective width at this location.

[0126] ,

[0127] in, The first in the continuous response diagram of human body shape The shape value of each pixel.

[0128] This invention characterizes the continuous extension of a structure in the vertical direction by accumulating pixel enhancement values ​​within a vertical window, while simultaneously calculating the effective width in a horizontal window. A normalized mapping is then performed using the effective width difference between adjacent rows to obtain a width stability coefficient. This allows the morphological response value to simultaneously reflect the stability of both vertical continuity and horizontal width variation. This process effectively reduces the interference of isolated local responses and transient noise on the results, enabling human structures with continuous morphological features and relatively gentle width variations to achieve higher response intensity.

[0129] In this embodiment, when obtaining morphological values Afterwards, through Map all morphological values ​​to the range of 0~255, where, For the first The mapping shape value of each pixel This is the maximum morphological value.

[0130] In this embodiment, the process of obtaining the significant structural response map includes:

[0131] Using three scale structuring elements, morphological top-hat transformations were performed on the structure separation map to obtain three scale top-hat response maps;

[0132] The global pixel mean is obtained by averaging the pixel values ​​of all pixels in the scale top cap response map.

[0133] In each scale top-hat response map, when the pixel value of a pixel is greater than the corresponding global pixel mean, the corresponding pixel is marked as a salient point;

[0134] In the three scale top-hat response maps, count the number of times the same pixel location is marked as a salient point;

[0135] Normalize the frequency of salient points to obtain the salient coefficient of each pixel;

[0136] At the same pixel point, the saliency coefficient is multiplied by the pixel value in the structure separation map to obtain the saliency structure response map.

[0137] In this embodiment, ,in, For the first Zhang's standard top hat response diagram. This is a structural separation diagram with noise constraints. For opening operation, For the first For each scale structuring element, m takes the values ​​1, 2, or 3. These correspond to the small-scale top-hat response map, the medium-scale top-hat response map, and the large-scale top-hat response map, respectively. These correspond to small-scale 3×3 rectangles, medium-scale 5×5 rectangles, and large-scale 7×7 rectangles, respectively.

[0138] In this embodiment, the frequency of salient points is normalized, i.e., the salient coefficient = frequency of salient points / 3.

[0139] This invention extracts local structural information at different scales through multi-scale morphological top-hat transformation, and uses the global pixel mean as a unified benchmark to screen salient points in the responses at each scale. Then, it generates normalized significance coefficients through cross-scale salient point statistics, giving higher weights to pixels with stable structural features across multiple scales. This process reduces the influence of single-scale noise or accidental enhancement regions on the results, highlighting regions with stable local structural features in the response map while preserving the spatial information of the original structure separation map.

[0140] The dual-channel neural network includes: a first multi-convolutional channel, a second multi-convolutional channel, a cross-cooperative weighted fusion unit, a deep feature extraction unit, a multi-layer spatial attention unit, and an output unit.

[0141] like Figure 2 As shown, the processing steps of a dual-channel neural network include:

[0142] The salient structure response map is processed using the first multi-convolutional channel to obtain the salient structure feature map;

[0143] The human body morphology continuous response map is processed using a second multi-convolutional channel to obtain the human body morphology feature map;

[0144] A cross-cooperative weighted fusion unit is used to fuse human morphological feature maps and salient structural feature maps to obtain a cooperative fusion feature map;

[0145] A deep feature extraction unit is used to extract deep features from the collaborative fusion feature map to obtain the collaborative fusion deep feature map;

[0146] Multi-layer spatial attention is applied to the collaboratively fused deep feature map using multi-layer spatial attention units to obtain a spatially enhanced feature map;

[0147] The output unit outputs human morphology recognition results based on spatially enhanced feature maps.

[0148] This invention employs a dual-channel neural network to independently extract features from the salient structure response map and the continuous response map of human morphology. The two types of features are complementarized and enhanced through a cross-cooperative weighted fusion unit. Then, a deep feature extraction unit is used to mine deep correlation features, and a multi-layer spatial attention unit is used to focus on the effective area and suppress background interference. Finally, the output unit outputs the recognition result, which improves the ability to distinguish human targets from interfering targets such as swaying trees and piles of debris in complex monitoring environments.

[0149] like Figure 3 As shown, the cross-cooperative weighted fusion unit includes: a first channel attention module, a second channel attention module, a multiplier M1, a multiplier M2, and an adder A1;

[0150] The first channel attention module is used to generate first channel attention based on the salient structural feature map;

[0151] The second-channel attention module is used to generate second-channel attention based on the human morphological feature map;

[0152] Multiplier M1 is used to multiply the first channel attention with the human body shape feature map element-wise to obtain the human body shape channel enhanced feature map;

[0153] Multiplier M2 is used to element-wise multiply the second channel attention with the salient structure feature map to obtain the salient structure channel enhanced feature map;

[0154] Adder A1 is used to add the enhanced feature map of the human body shape channel and the enhanced feature map of the salient structure channel element by element to obtain a synergistic fusion feature map.

[0155] The cross-cooperative weighted fusion unit generates corresponding attention through the first channel attention module and the second channel attention module, respectively. It uses multipliers to enhance the channels of human morphological feature maps and salient structural feature maps, and then uses adders to fuse the two types of enhanced feature maps. This enables mutual guidance and complementary enhancement between the two types of features, making human morphological features and salient structural features accurately matched and effectively enhanced in the channel dimension, avoiding feature interference and information loss, and improving the representation ability and discriminability of the fused features.

[0156] In this embodiment, the first channel attention module and the second channel attention module have the same structure, both including a 1×1 convolutional layer and a sigmoid layer connected in sequence.

[0157] like Figure 4 As shown, the multi-layer spatial attention unit includes: a first spatial attention module, a second spatial attention module, a multiplier M3, a multiplier M4, and an adder A2;

[0158] The first spatial attention module is used to generate first spatial attention on the collaboratively fused deep feature map;

[0159] The second spatial attention module is used to generate second spatial attention from the collaboratively fused deep feature map;

[0160] Multiplier M3 is used to element-wise multiply the first spatial attention with the collaborative fusion deep feature map to obtain the first spatial augmentation map;

[0161] Multiplier M4 is used to element-wise multiply the second spatial attention with the collaborative fusion deep feature map to obtain the second spatial augmentation map;

[0162] Adder A2 is used to add the first spatial augmentation map and the second spatial augmentation map element by element to obtain a spatial augmentation feature map.

[0163] In this embodiment, the first spatial attention module includes a 3×3 convolutional layer and a sigmoid layer connected in sequence; the second spatial attention module includes a 7×7 convolutional layer and a sigmoid layer connected in sequence.

[0164] The multi-layer spatial attention unit generates different spatial attention on the collaboratively fused deep feature map through the first spatial attention module and the second spatial attention module respectively. Then, it uses a multiplier to perform spatial weighted enhancement with the collaboratively fused deep feature map. Finally, it adds the two spatially enhanced feature maps with an adder to obtain a spatially enhanced feature map. This can adaptively focus on the human target region from multiple spatial scales and enhance the spatial expression of effective features.

[0165] like Figure 5 As shown, the first and second multi-convolutional channels have the same structure, both including a first convolutional layer, a second convolutional layer and a third convolutional layer connected in sequence. The kernel size of the first convolutional layer is 3×3, the kernel size of the second convolutional layer is 5×5, and the kernel size of the third convolutional layer is 3×3.

[0166] like Figure 6 As shown, the deep feature extraction unit includes: a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, and an adder A3. The kernel size of the fourth, fifth, and sixth convolutional layers is 3×3. The adder A3 is used for residual connections.

[0167] The output unit uses a fully connected layer, takes the spatial augmented feature map as input for binary classification, and outputs either "human target" or "non-human target". Example 2: An intrusion prevention monitoring and early warning system includes: a filtering subsystem, a directional augmentation weight acquisition subsystem, a first construction subsystem, a second construction subsystem, a third construction subsystem, and an identification and early warning subsystem;

[0168] The filtering subsystem is used to perform Gaussian filtering with a noise threshold on the motion difference image to obtain a structure separation map;

[0169] The orientation enhancement weight acquisition subsystem is used to extract and filter the contour of the structure separation map and obtain the orientation enhancement weight of each pixel.

[0170] The first construction subsystem is used to multiply the pixel value of each pixel in the structure separation map with the orientation enhancement weight of the corresponding pixel to obtain the pixel enhancement value and construct the orientation enhancement map;

[0171] The second construction subsystem is used to multiply the longitudinal continuity intensity of the orientation enhancement map with the width stability coefficient to construct a continuous response map of human body shape.

[0172] The third construction subsystem is used to perform multi-scale morphological top-hat transformation on the structure separation map. By statistically analyzing cross-scale salient points, it obtains significance coefficients, modulates the structure separation map, and constructs a salient structure response map.

[0173] The identification and early warning subsystem uses a dual-channel neural network to process the salient structure response map and the continuous response map of human body shape to obtain the human body shape recognition result. When the human body shape recognition result is a human body, an intrusion warning is issued.

[0174] The specific implementation process of Example 2 is the same as that of Example 1.

[0175] Since traditional machine learning classifiers have limited feature recognition capabilities, this invention constructs a continuous response map of human body shape and a salient structure response map, which makes human features more salient and reduces the probability of misjudgment by the neural network. At the same time, a dual-channel neural network is constructed to process the salient structure response map and the continuous response map of human body shape, thereby improving the accuracy of intrusion monitoring and early warning.

[0176] This invention replaces traditional median filtering with Gaussian filtering that includes a noise threshold, accurately separating noise from structural points to preserve the integrity of the human body's edge contour. It then extracts and filters contours from the structural separation map and obtains directional enhancement weights to construct a directional enhancement map. Furthermore, it fuses longitudinal continuous intensity and width stability coefficients to construct a continuous response map of the human body's morphology. Simultaneously, it constructs a salient structural response map through multi-scale morphological top-hat transformation and cross-scale salient point statistics. Finally, it employs a dual-channel neural network to process the dual response maps, fusing global morphology and local salient features, effectively improving the ability to distinguish between interfering targets and human targets, reducing false positives and false negatives, thereby solving the problem of low accuracy in existing intrusion prevention monitoring and early warning technologies.

[0177] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intrusion monitoring and early warning method, characterized in that, include: The motion difference image is subjected to Gaussian filtering with a noise threshold to obtain a structure separation map. Contour extraction and filtering are performed on the structure separation map to obtain the orientation enhancement weight of each pixel; The pixel value of each pixel in the structure separation map is multiplied by the orientation enhancement weight of the corresponding pixel to obtain the pixel enhancement value, and the orientation enhancement map is constructed. The longitudinal continuity intensity of the directional enhancement map is multiplied by the width stability coefficient to construct a continuous response map of human body shape. Multi-scale morphological top-hat transformation is performed on the structure separation map. Significance coefficients are obtained through cross-scale salient point statistics. The structure separation map is then modulated to construct a salient structure response map. A dual-channel neural network is used to process the salient structure response map and the human body shape continuous response map to obtain the human body shape recognition result. When the human body shape recognition result is a human body, an intrusion warning is issued. The Gaussian filtering process includes: A Gaussian filter is applied to the motion difference image to obtain a smoothed image; The pixel difference image is obtained by subtracting the smoothed image from the motion difference image pixel by pixel and taking the absolute value. In a pixel difference image, when the pixel value is less than the noise threshold, the corresponding pixel is considered an interference point. In a pixel difference image, when the pixel value is greater than or equal to the noise threshold, the corresponding pixel is a structure point; For interference points, the pixel values ​​of the smoothed image are used as the output; for structural points, the pixel values ​​of the motion difference image are used as the output, thus obtaining the structural separation map. The process of obtaining the orientation enhancement weights for each pixel includes: Canny edge detection is performed on the structure separation map to obtain multiple edge contours; Calculate the length of each edge contour, filter out edge contours whose length is greater than or equal to the minimum human body contour length threshold, and form a candidate contour set. For each edge contour in the candidate contour set, calculate the average gradient direction of all pixels inside it to obtain the average gradient direction of the corresponding edge contour. The orientation enhancement weight of each pixel is obtained based on the difference between the gradient direction of each pixel in each edge contour in the candidate contour set and the average gradient direction of the corresponding edge contour. The process of obtaining the orientation enhancement weights for each pixel includes: The gradient direction of each pixel and its neighboring pixels in each edge contour of the candidate contour set is subtracted from the average gradient direction, and the direction deviation is obtained by using an exponential decay term. Convert the gradient directions of a pixel and its neighboring pixels into complex numbers. The gradient direction in complex form is multiplied by the deviation from the direction, and the average value is taken to obtain the average complex vector; The magnitude of the average complex vector is used as the orientation enhancement weight for the corresponding pixel, and the orientation enhancement weights for other pixels that do not belong to the candidate contour set are reset to 0. The process of constructing a continuous response map of human body shape includes: Using each pixel in the directional enhancement map as the center, the pixel enhancement values ​​of all pixels within the vertical window are added together to obtain the vertical continuous intensity of the corresponding pixel. Centered on each pixel in the orientation enhancement map, within a horizontal window, count the number of pixels whose pixel enhancement value is greater than the average pixel enhancement value, and use this number of pixels as the effective width of the corresponding pixel. The width stability coefficient of each pixel is obtained by comparing the effective width of each pixel with the pixels in the adjacent row. Multiply the vertical continuous intensity of the same pixel by the width stability coefficient to obtain the shape value of the corresponding pixel and construct a continuous response map of human body shape. The process of obtaining a significant structural response map includes: Using three scale structuring elements, morphological top-hat transformations were performed on the structure separation map to obtain three scale top-hat response maps; The global pixel mean is obtained by averaging the pixel values ​​of all pixels in the scale top cap response map. In each scale top-hat response map, when the pixel value of a pixel is greater than the corresponding global pixel mean, the corresponding pixel is marked as a salient point; In the three scale top-hat response maps, count the number of times the same pixel location is marked as a salient point; Normalize the frequency of salient points to obtain the salient coefficient of each pixel; At the same pixel point, the saliency coefficient is multiplied by the pixel value in the structure separation map to obtain the saliency structure response map.

2. The intrusion monitoring and warning method according to claim 1, wherein The processing steps of a two-channel neural network include: The salient structure response map is processed using the first multi-convolutional channel to obtain the salient structure feature map; The human body morphology continuous response map is processed using a second multi-convolutional channel to obtain the human body morphology feature map; A cross-cooperative weighted fusion unit is used to fuse human morphological feature maps and salient structural feature maps to obtain a cooperative fusion feature map; A deep feature extraction unit is used to extract deep features from the collaborative fusion feature map to obtain the collaborative fusion deep feature map; Multi-layer spatial attention is applied to the collaboratively fused deep feature map using multi-layer spatial attention units to obtain a spatially enhanced feature map; The output unit outputs human morphology recognition results based on spatially enhanced feature maps.

3. The intrusion monitoring and warning method according to claim 2, wherein, The cross-cooperative weighted fusion unit includes: a first-channel attention module, a second-channel attention module, multiplier M1, multiplier M2, and adder A1; The first channel attention module is used to generate first channel attention based on the salient structural feature map; The second-channel attention module is used to generate second-channel attention based on the human morphological feature map; Multiplier M1 is used to multiply the first channel attention with the human body shape feature map element-wise to obtain the human body shape channel enhanced feature map; Multiplier M2 is used to element-wise multiply the second channel attention with the salient structure feature map to obtain the salient structure channel enhanced feature map; Adder A1 is used to add the enhanced feature map of the human body shape channel and the enhanced feature map of the salient structure channel element by element to obtain a synergistic fusion feature map.

4. The intrusion monitoring and warning method according to claim 2, wherein The multi-layer spatial attention unit includes: a first spatial attention module, a second spatial attention module, a multiplier M3, a multiplier M4, and an adder A2; The first spatial attention module is used to generate first spatial attention on the collaboratively fused deep feature map; The second spatial attention module is used to generate second spatial attention from the collaboratively fused deep feature map; Multiplier M3 is used to element-wise multiply the first spatial attention with the collaborative fusion deep feature map to obtain the first spatial augmentation map; Multiplier M4 is used to element-wise multiply the second spatial attention with the collaborative fusion deep feature map to obtain the second spatial augmentation map; Adder A2 is used to add the first spatial augmentation map and the second spatial augmentation map element by element to obtain a spatial augmentation feature map.

5. An intrusion prevention monitoring and early warning system, implemented based on the intrusion prevention monitoring and early warning method according to any one of claims 1 to 4, characterized in that, include: The system comprises a filtering subsystem, a direction enhancement weight acquisition subsystem, a first construction subsystem, a second construction subsystem, a third construction subsystem, and a recognition and early warning subsystem. The filtering subsystem is used to perform Gaussian filtering with a noise threshold on the motion difference image to obtain a structure separation map; The orientation enhancement weight acquisition subsystem is used to extract and filter the contour of the structure separation map and obtain the orientation enhancement weight of each pixel. The first construction subsystem is used to multiply the pixel value of each pixel in the structure separation map with the orientation enhancement weight of the corresponding pixel to obtain the pixel enhancement value and construct the orientation enhancement map; The second construction subsystem is used to multiply the longitudinal continuity intensity of the orientation enhancement map with the width stability coefficient to construct a continuous response map of human body shape. The third construction subsystem is used to perform multi-scale morphological top-hat transformation on the structure separation map. By statistically analyzing cross-scale salient points, it obtains significance coefficients, modulates the structure separation map, and constructs a salient structure response map. The identification and early warning subsystem uses a dual-channel neural network to process the salient structure response map and the continuous response map of human morphology to obtain the human morphology recognition result. When the human morphology recognition result is a human body, an intrusion warning is issued.

Images