House safety monitoring platform based on unmanned aerial vehicle detection

Abstract

The invention relates to the technical field of house safety monitoring, in particular to a house safety monitoring platform based on unmanned aerial vehicle detection, which comprises an unmanned aerial vehicle carrying vehicle, a detection unmanned aerial vehicle, a data processing module and a data analysis module, and is characterized in that the unmanned aerial vehicle carrying vehicle is used for carrying and supplementing energy for the detection unmanned aerial vehicle; the detection unmanned aerial vehicle is provided with an acquisition sensor module and a point marking spray gun; according to the house safety monitoring platform, the unmanned aerial vehicle carries the acquisition sensor to detect the safety condition of the outer facade of the house, an area which is difficult to reach by workers can be covered, detection of a large number of house facades can be completed in a short time, the detection efficiency is improved, and the risk of manual climbing is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of house safety monitoring, and in particular to a house safety monitoring platform based on drone detection. Background Art

[0002] Common problems with the exterior facade of a house include cracks, hollowing, and exterior wall shedding, etc., which are mostly found in medium and high-rise buildings. As the service life increases, the problems of exterior wall insulation increase year by year. Therefore, problems such as cracking, shedding, and hollowing of exterior wall insulation are common, causing life and property safety. Therefore, regular inspection of the facade is an important measure to ensure the safety of the building and extend its service life. Through professional inspections, the current status of the facade can be evaluated, and potential safety hazards such as cracks and shedding can be discovered in a timely manner. The existing facade inspection is mainly based on manual inspections, which has low work efficiency, and the manual climbing work scene is relatively dangerous, and the inspection cycle is long. Summary of the invention

[0003] The purpose of the present invention is to provide a housing safety monitoring platform based on drone detection to solve the problems raised in the above background technology.

[0004] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a house safety monitoring platform based on drone detection, comprising a drone carrying vehicle, a detection drone, a data processing module and a data analysis module, wherein the drone carrying vehicle is used to carry and replenish the detection drone, and the detection drone is provided with a collection sensor module and a punctuation spray gun, and the detection drone can spray marks on the outer surface of the house wall through the punctuation spray gun, and the detection drone collects data on the outer surface of the house wall through the collection sensor module, and the collected data is input into the data processing module for preprocessing, and then input into the data analysis module to calculate the house safety status through an algorithm.

[0005] The acquisition sensor module includes a high-definition camera, an infrared camera and a radar. The data analysis module is provided with an image detection algorithm with an automatic optimization function. The image detection algorithm detects cracking and falling off of the outer surface of the house wall according to the image data collected by the high-definition camera. The infrared camera detects thermal defects and leakage of the house wall through thermal imaging technology. The radar is used for hollowing detection. When the radar detects hollowing, the outer surface of the hollow position of the house wall is marked by spraying with a punctuation spray gun. The data analysis module is provided with a combination algorithm. The combination algorithm calculates the hollowing separation distance of the marked point by combining the image data and the radar acquisition data, so as to realize long-term monitoring of fixed reference points for hollowing separation.

[0006] The image detection algorithm includes: using the Canny edge detection algorithm to extract edge information in the house wall image, calculating the gradient amplitude and direction of the image, and the gradient amplitude calculation formula is: Among them, G x and G y are the gradient values ​​of the image in the x and y directions, respectively, which can be obtained by calculating the first-order partial derivative of the image; non-maximum suppression, by comparing the gradient amplitude of the current pixel with the gradient amplitude of the adjacent pixel points along its gradient direction, retaining the local maximum point as the edge candidate point; dual threshold detection and connection, setting a high threshold and a low threshold, points above the high threshold are determined as edge points, points below the low threshold are discarded, and points between the two are also determined as edge points if they are connected to the determined edge points;

[0007] Gray-level co-occurrence matrix GLCM is used to extract the texture features of the image. For a given image I, at a certain distance d and direction θ, the gray-level co-occurrence matrix P(i, j, d, θ) represents the probability that a pixel with gray value i and a pixel with gray value j in the image appear at the same time at a distance d and direction θ. The texture features of the image are described by calculating the statistics of GLCM, including contrast, correlation, energy and entropy.

[0008] According to the results of edge detection and texture feature extraction, set a suitable threshold to segment the image. For the edge intensity value, set a threshold T edge , the area with edge strength higher than the threshold is preliminarily judged as the area where there may be peeling or cracking. For texture features, taking contrast as an example, the threshold T is set contrast , when the contrast of a certain area is higher than the threshold, it is also included in the scope of suspicious areas;

[0009] Morphological operations are performed on the suspicious areas that are initially segmented, including corrosion and expansion. The corrosion operation can remove some isolated small points or noise, and the expansion operation can fill some holes or connect adjacent areas to make the suspicious areas more complete and continuous. The basic corrosion operation formula is as follows: Where A is the original image, B is the structural element, (B) x Indicates that the structural element B is translated to point x; the basic operation formula of expansion is as follows:

[0010] Establish a feature template library of known house wall peeling and cracking areas, match the features of the segmented suspicious areas, including edge features and texture features, with the templates in the feature template library, and use the distance measurement method Euclidean distance to calculate the matching degree. The calculation formula is: Among them, x i is the i-th element of the feature vector of the suspicious area, y iIt is the i-th element of the feature vector of a template in the feature template library. It determines whether the suspicious area is really a detached or cracked area according to the degree of matching;

[0011] Set a matching threshold T match When the matching degree between the suspicious area and all templates in the feature template library is lower than the threshold, the area is marked as an uncertain area, and the image data and related feature information of the uncertain area are presented to the manual judge through the system interface. The manual judge determines whether there is detachment or cracking in the area according to the situation, and feeds back the judgment result to the algorithm system, which performs automatic optimization learning.

[0012] The automatic optimization learning specifically includes: when it is manually determined that there is a detached or cracked area, and the features of the area are greatly different from the existing template features, consider adding the features of the area to the feature template library as a new template; first, it is necessary to determine the measurement standard of feature differences. Suppose the Euclidean distance is used to measure the difference between the feature vector of the suspicious area and the feature vector of the existing template, and the feature vector of the suspicious area is There are m templates in the existing template library, and the feature vector of the jth template is Then the Euclidean distance calculation formula between the suspicious area and the jth template is: Set a distance threshold T new , if for all j = 1, 2, ..., m, there is d j >T new , it is considered that the suspicious area features are greatly different from the existing template features, and it can be added to the feature template library as a new template;

[0013] When an area is manually determined to have fallen off or cracked, and its features are different from the existing template but not to the extent of adding a new template, the existing template is adjusted and optimized; the weighted average method is used to update the template features. The feature vector of the suspicious area manually determined to have problems is The most similar existing template feature vector is Its Euclidean distance is d min ; Define a learning rate α to control the update amplitude, the updated template feature vector The calculation formulas for each element of are as follows:

[0014] For areas that are manually judged as not falling off or cracking, but were previously misjudged as suspicious areas, it is necessary to adjust the model's discrimination threshold and other factors to reduce the misjudgment rate. Assume that the Euclidean distance was used for judgment based on feature matching, and the matching threshold T was set. matchTo determine whether the suspicious area is actually a peeling or cracking area; count the number of times N that it is misjudged as a problematic area in a certain number of samples. falsepositive And the total number of samples N total , calculate the misjudgment rate If the false positive rate is higher than an acceptable threshold T fp , then the matching threshold T match Adjust the threshold proportionally and set the adjusted matching threshold to Among them, β is an adjustment coefficient used to control the amplitude of threshold adjustment.

[0015] If there are areas that are manually judged to have fallen off or cracked, but were previously misjudged as non-problematic areas, adjustments should also be made, and the number of times the areas were misjudged as non-problematic areas should be counted. And the total number of samples N total , calculate the misjudgment rate If the misjudgment rate is higher than the acceptable threshold T fn , it is necessary to adjust the parameters of the feature extraction link. Taking the edge detection threshold adjustment as an example, let the current edge detection threshold be T edg , the adjusted edge detection threshold is T e ' dg , the following formula can be used: Among them, γ is an adjustment coefficient used to control the amplitude of threshold adjustment. The same idea can be applied to the adjustment of other related parameters.

[0016] The combined algorithm includes: establishing a two-dimensional rectangular coordinate system on the image plane, taking a corner of the image as the origin, horizontally to the right as the positive direction of the x-axis, and vertically downward as the positive direction of the y-axis, and determining the coordinates (x i ,y i );

[0017] Through the system calibration performed in advance, the conversion relationship between the radar scanning space and the image plane coordinate system is established. By placing multiple reference points at known positions, their radar data and position information in the image are obtained at the same time, and then a mathematical model is used to fit the conversion relationship between the two; assuming that the conversion relationship from image coordinates to radar scanning space coordinates obtained after calibration is: Where (x r ,y r ,z r ) are the coordinates of the fixed smear mark point in the radar scanning space, and a1, b1, c1, a2, b2, c2, a3, b3, c3 are the conversion coefficients obtained through calibration.

[0018] Based on the preprocessed radar data, an algorithm is used to identify the hollow areas in the wall. A reasonable reflection wave intensity threshold T is set. When the intensity of the radar echo signal in a certain area is lower than T, the area is considered to be a hollow area. The boundary extraction and shape fitting of the identified hollow area are performed to obtain the geometric shape description of the hollow area. A polygon can be used to approximate the hollow area, and the coordinate values ​​of its vertex coordinates in the radar scanning space are recorded.

[0019] Assume that the coordinates of a vertex in the hollow drum area are (x g ,y g ,z g ), the coordinates of the fixed smear mark point in the radar scanning space are (x r ,y r ,z r ), then the distance d from the fixed mark point to the vertex of the hollow area g It can be calculated by the distance formula: Monitoring and tracking in multiple detections g of changes.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] The housing safety monitoring platform of the present invention detects the safety of the facade of the house by using a drone equipped with a collection sensor. It can cover areas that are difficult to reach manually and can complete the detection of a large number of house facades in a short time, thereby improving the detection efficiency and reducing the risk of manual climbing.

[0022] The present invention can detect cracking and falling off of the outer surface of the house wall according to the image data collected by the high-definition camera through the set image detection algorithm, and has an automatic optimization and adjustment function, which can gradually improve the recognition accuracy during long-term use.

[0023] The present invention can mark the hollow area with spray points by means of a punctuation spray gun, combined with algorithms and radar modules, identify the spray point positioning point through a high-definition camera, and calculate the separation distance of the point through the radar combined with the algorithm. Therefore, when the hollow area cannot be repaired in a short time, the change of the hollow separation distance of the point can be tracked in multiple detections to achieve safe monitoring of the hollow area. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a module block diagram of the housing safety monitoring platform of the present invention. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] See also Figure 1 The present invention provides a technical solution: a housing safety monitoring platform based on drone detection, including a drone carrier vehicle, a detection drone, a data processing module and a data analysis module. The drone carrier vehicle is used to carry and recharge the detection drone. The detection drone takes off from the drone carrier vehicle and is charged by the drone carrier vehicle. The detection drone is provided with a collection sensor module and a punctuation spray gun. The detection drone can spray marks on the outer surface of the house wall through the punctuation spray gun. The punctuation spray gun is a device that can spray marking pigments at a long distance in the prior art. The collection sensor module includes a high-definition camera, an infrared camera and a radar. The detection drone collects data on the outer surface of the house wall through the collection sensor module. The collected data is input into the data processing module for preprocessing, and then input into the data analysis module to calculate the housing safety situation through an algorithm. The preprocessing of the collected data in the data processing module includes graying, filtering and denoising, and image enhancement for image data. For radar data, it includes filtering the radar echo signal to remove noise interference and improve the clarity and recognizability of the signal. Common filtering algorithms, such as mean filtering and median filtering, can be used to parse the filtered radar data and convert it into depth information data in units of distance according to the radar's scanning parameters, such as scanning angle and resolution.

[0027] Infrared cameras use thermal imaging technology to detect thermal defects and leaks in the walls of houses. That is, the appearance of thermal defects and leakage points in the walls will cause a significant temperature difference in that area relative to the entire wall. Once areas with large temperature differences are detected, a judgment can be made. At the same time, the high-definition camera can be combined with the recognition of objects to automatically exclude equipment such as air conditioner outdoor units.

[0028] The data analysis module is equipped with an image detection algorithm with automatic optimization function. The image detection algorithm detects the cracking and peeling of the outer surface of the house wall based on the image data collected by the high-definition camera. The image detection algorithm includes: using the Canny edge detection algorithm to extract the edge information in the house wall image, calculating the gradient amplitude and direction of the image, and the gradient amplitude calculation formula is: Among them, G x and G yare the gradient values ​​of the image in the x and y directions, respectively, which can be obtained by calculating the first-order partial derivative of the image; non-maximum suppression, by comparing the gradient amplitude of the current pixel with the gradient amplitude of the adjacent pixel points along its gradient direction, retaining the local maximum point as the edge candidate point; dual threshold detection and connection, setting a high threshold and a low threshold, points above the high threshold are determined as edge points, points below the low threshold are discarded, and points between the two are also determined as edge points if they are connected to the determined edge points;

[0029] The gray-level co-occurrence matrix GLCM is used to extract the texture features of the image. For a given image I, at a certain distance d and direction θ, the gray-level co-occurrence matrix P(i, j, d, θ) represents the probability that the pixel with gray value i and the pixel with gray value j in the image appear at the same time at a distance d and direction θ. The texture features of the image are described by calculating the statistics of GLCM, including contrast, correlation, energy and entropy. Taking contrast as an example, the calculation formula is:

[0030] According to the results of edge detection and texture feature extraction, set a suitable threshold to segment the image. For the edge intensity value, set a threshold T edge , the area with edge strength higher than the threshold is preliminarily judged as the area where there may be peeling or cracking. For texture features, taking contrast as an example, the threshold T is set contrast , when the contrast of a certain area is higher than the threshold, it is also included in the scope of suspicious areas;

[0031] Morphological operations are performed on the suspicious areas that are initially segmented, including corrosion and expansion. The corrosion operation can remove some isolated small points or noise, and the expansion operation can fill some holes or connect adjacent areas to make the suspicious areas more complete and continuous. The basic corrosion operation formula is as follows: Where A is the original image, B is the structural element, (B) x Indicates that the structural element B is translated to point x; the basic operation formula of expansion is as follows:

[0032] Establish a feature template library of known house wall peeling and cracking areas, match the features of the segmented suspicious areas, including edge features and texture features, with the templates in the feature template library, and use the distance measurement method Euclidean distance to calculate the matching degree. The calculation formula is: Among them, x i is the i-th element of the feature vector of the suspicious area, y i It is the i-th element of the feature vector of a template in the feature template library. It determines whether the suspicious area is really a detached or cracked area according to the degree of matching;

[0033] Set a matching threshold T match When the matching degree between the suspicious area and all templates in the feature template library is lower than the threshold, the area is marked as an uncertain area, and the image data and related feature information of the uncertain area are presented to the manual judge through the system interface. The manual judge determines whether there is detachment or cracking in the area according to the situation, and feeds back the judgment result to the algorithm system, which performs automatic optimization learning.

[0034] Automatic optimization learning specifically includes: when an area is manually determined to have fallen off or cracked, and the features of the area are significantly different from the existing template features, the features of the area are considered to be added to the feature template library as a new template; first, the measurement standard of feature differences needs to be determined. Suppose the Euclidean distance is used to measure the difference between the feature vector of the suspicious area and the feature vector of the existing template. Suppose the feature vector of the suspicious area is There are m templates in the existing template library, and the feature vector of the jth template is Then the Euclidean distance calculation formula between the suspicious area and the jth template is: Set a distance threshold T new , if for all j = 1, 2, ..., m, there is d j >T new , it is considered that the suspicious area features are greatly different from the existing template features, and it can be added to the feature template library as a new template;

[0035] When an area is manually determined to have fallen off or cracked, and its features are different from the existing template but not to the extent of adding a new template, the existing template is adjusted and optimized; the weighted average method is used to update the template features. The feature vector of the suspicious area manually determined to have problems is The most similar existing template feature vector is Its Euclidean distance is d min ; Define a learning rate α to control the update amplitude, the updated template feature vector The calculation formulas for each element of are as follows:

[0036] For areas that are manually judged as not falling off or cracking, but were previously misjudged as suspicious areas, it is necessary to adjust the model's discrimination threshold and other factors to reduce the misjudgment rate. Assume that the Euclidean distance was used for judgment based on feature matching, and the matching threshold T was set. match To determine whether the suspicious area is actually a peeling or cracking area; count the number of times N that it is misjudged as a problematic area in a certain number of samples. falsepositive And the total number of samples N total , calculate the misjudgment rate If the false positive rate is higher than an acceptable threshold T fp , then the matching threshold T match Adjust the threshold proportionally and set the adjusted matching threshold to Among them, β is an adjustment coefficient used to control the amplitude of threshold adjustment.

[0037] If there are areas that are manually judged to have fallen off or cracked, but were previously misjudged as non-problematic areas, adjustments should also be made, and the number of times the areas were misjudged as non-problematic areas should be counted. And the total number of samples N total , calculate the misjudgment rate If the misjudgment rate is higher than the acceptable threshold T fn , it is necessary to adjust the parameters of the feature extraction link. Taking the edge detection threshold adjustment as an example, let the current edge detection threshold be T edg , the adjusted edge detection threshold is T e ' dg , the following formula can be used: T e ' dge =T edge +γ×(P falsenegative -T fn ), where γ is an adjustment coefficient used to control the amplitude of threshold adjustment. The same idea can be applied to the adjustment of other related parameters.

[0038] The radar is used for hollowing detection. When the radar detects hollowing, the outer surface of the hollowing position of the house wall is sprayed with a punctuation spray gun. A combination algorithm is set in the data analysis module. The combination algorithm combines the image data and the radar acquisition data to calculate the hollowing separation distance of the marked point to achieve long-term monitoring of the fixed reference point of the hollowing separation. The combination algorithm includes: establishing a two-dimensional rectangular coordinate system on the image plane, with a corner of the image as the origin, the horizontal right as the positive direction of the x-axis, and the vertical downward as the positive direction of the y-axis, and determining the coordinates (x i ,y i );

[0039] Through the system calibration performed in advance, the conversion relationship between the radar scanning space and the image plane coordinate system is established. By placing multiple reference points at known positions, their radar data and position information in the image are obtained at the same time, and then a mathematical model is used to fit the conversion relationship between the two; assuming that the conversion relationship from image coordinates to radar scanning space coordinates obtained after calibration is: Where (x r ,y r ,z r) are the coordinates of the fixed smear mark point in the radar scanning space, and a1, b1, c1, a2, b2, c2, a3, b3, c3 are the conversion coefficients obtained through calibration.

[0040] Based on the preprocessed radar data, an algorithm is used to identify the hollow areas in the wall. A reasonable reflection wave intensity threshold T is set. When the intensity of the radar echo signal in a certain area is lower than T, the area is considered to be a hollow area. The boundary extraction and shape fitting of the identified hollow area are performed to obtain the geometric shape description of the hollow area. A polygon can be used to approximate the hollow area, and the coordinate values ​​of its vertex coordinates in the radar scanning space are recorded.

[0041] Assume that the coordinates of a vertex in the hollow drum area are (x g ,y g ,z g ), the coordinates of the fixed smear mark point in the radar scanning space are (x r ,y r ,z r ), then the distance d from the fixed mark point to the vertex of the hollow area g It can be calculated by the distance formula: Monitoring and tracking in multiple detections g of changes.

[0042] That is, when the radar detects hollowing, the outer surface of the house wall at the hollowing position is sprayed with a punctuation spray gun, and then the spray mark is recognized by the high-definition camera, and the coordinate value mapped in the radar scanning space is calculated by the algorithm to calculate the separation distance between the hollow layer and the wall at the spray mark. This is a one-time detection. When the hollowing cannot be repaired in a short time, it can be detected multiple times regularly, such as once a day. g The value of d g Whether the value continues to increase, determine whether the degree of hollow separation has a tendency to increase, and achieve long-term safety monitoring.

[0043] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A housing safety monitoring platform based on drone detection, comprising a drone carrier vehicle, a detection drone, a data processing module and a data analysis module, characterized in that: The drone carrier vehicle is used to carry and recharge the inspection drone. The inspection drone is provided with a collection sensor module and a punctuation spray gun. The inspection drone can spray marks on the outer surface of the house wall through the punctuation spray gun. The inspection drone collects data on the outer surface of the house wall through the collection sensor module. The collected data is input into the data processing module for preprocessing, and then input into the data analysis module to calculate the safety status of the house through an algorithm.

2. A housing safety monitoring platform based on drone detection according to claim 1, characterized in that: The acquisition sensor module includes a high-definition camera, an infrared camera and a radar. The data analysis module is provided with an image detection algorithm with an automatic optimization function. The image detection algorithm detects cracking and falling off of the outer surface of the house wall according to the image data collected by the high-definition camera. The infrared camera detects thermal defects and leakage of the house wall through thermal imaging technology. The radar is used for hollowing detection. When the radar detects hollowing, the outer surface of the hollow position of the house wall is marked by spraying with a punctuation spray gun. The data analysis module is provided with a combination algorithm. The combination algorithm calculates the hollowing separation distance of the marked point by combining the image data and the radar acquisition data, so as to realize long-term monitoring of fixed reference points for hollowing separation.

3. A housing safety monitoring platform based on drone detection according to claim 2, characterized in that: The image detection algorithm includes: using the Canny edge detection algorithm to extract edge information in the house wall image, calculating the gradient amplitude and direction of the image, and the gradient amplitude calculation formula is: Among them, G x and G y are the gradient values ​​of the image in the x and y directions, respectively, which can be obtained by calculating the first-order partial derivative of the image; non-maximum suppression, by comparing the gradient amplitude of the current pixel with the gradient amplitude of the adjacent pixel points along its gradient direction, retaining the local maximum point as the edge candidate point; dual threshold detection and connection, setting a high threshold and a low threshold, points above the high threshold are determined as edge points, points below the low threshold are discarded, and points between the two are also determined as edge points if they are connected to the determined edge points; Gray-level co-occurrence matrix GLCM is used to extract the texture features of the image. For a given image I, at a certain distance d and direction θ, the gray-level co-occurrence matrix P(i, j, d, θ) represents the probability that a pixel with gray value i and a pixel with gray value j in the image appear at the same time at a distance d and direction θ. The texture features of the image are described by calculating the statistics of GLCM, including contrast, correlation, energy and entropy. According to the results of edge detection and texture feature extraction, set a suitable threshold to segment the image. For the edge intensity value, set a threshold T edge , the area with edge strength higher than the threshold is preliminarily judged as the area where there may be peeling or cracking. For texture features, taking contrast as an example, the threshold T is set contrast , when the contrast of a certain area is higher than the threshold, it is also included in the scope of suspicious areas; Morphological operations are performed on the suspicious areas that are initially segmented, including corrosion and expansion. The corrosion operation can remove some isolated small points or noise, and the expansion operation can fill some holes or connect adjacent areas to make the suspicious areas more complete and continuous. The basic corrosion operation formula is as follows: Where A is the original image, B is the structural element, (B) x Indicates that the structural element B is translated to point x; the basic operation formula of expansion is as follows: Establish a feature template library of known house wall peeling and cracking areas, match the features of the segmented suspicious areas, including edge features and texture features, with the templates in the feature template library, and use the distance measurement method Euclidean distance to calculate the matching degree. The calculation formula is: Among them, x i is the i-th element of the feature vector of the suspicious area, y i It is the i-th element of the feature vector of a template in the feature template library. It determines whether the suspicious area is really a detached or cracked area according to the degree of matching; Set a matching threshold T match When the matching degree between the suspicious area and all templates in the feature template library is lower than the threshold, the area is marked as an uncertain area, and the image data and related feature information of the uncertain area are presented to the manual judge through the system interface. The manual judge determines whether there is detachment or cracking in the area according to the situation, and feeds back the judgment result to the algorithm system, which performs automatic optimization learning.

4. A housing safety monitoring platform based on drone detection according to claim 3, characterized in that: The automatic optimization learning specifically includes: when it is manually determined that there is a detached or cracked area, and the features of the area are greatly different from the existing template features, consider adding the features of the area to the feature template library as a new template; first, it is necessary to determine the measurement standard of feature differences. Suppose the Euclidean distance is used to measure the difference between the feature vector of the suspicious area and the feature vector of the existing template, and the feature vector of the suspicious area is There are m templates in the existing template library, and the feature vector of the jth template is Then the Euclidean distance calculation formula between the suspicious area and the jth template is: Set a distance threshold T new , if for all j = 1, 2, ..., m, there is d j >T new , it is considered that the suspicious area features are greatly different from the existing template features, and it can be added to the feature template library as a new template; When an area is manually determined to have fallen off or cracked, and its features are different from the existing template but not to the extent of adding a new template, the existing template is adjusted and optimized; the weighted average method is used to update the template features. The feature vector of the suspicious area manually determined to have problems is The most similar existing template feature vector is Its Euclidean distance is d min ; Define a learning rate α to control the update amplitude, the updated template feature vector The calculation formulas for each element of are as follows:

5. A housing safety monitoring platform based on drone detection according to claim 4, characterized in that: For areas that are manually judged as not falling off or cracking, but were previously misjudged as suspicious areas, it is necessary to adjust the model's discrimination threshold and other factors to reduce the misjudgment rate. Assume that the Euclidean distance was used for judgment based on feature matching, and the matching threshold T was set. match To determine whether the suspicious area is actually a peeling or cracking area; count the number of times N that it is misjudged as a problematic area in a certain number of samples. falsepositive And the total number of samples N total , calculate the misjudgment rate If the false positive rate is higher than an acceptable threshold T fp , then the matching threshold T match Adjust the threshold proportionally and set the adjusted matching threshold to Among them, β is an adjustment coefficient used to control the amplitude of threshold adjustment.

6. A housing safety monitoring platform based on drone detection according to claim 5, characterized in that: If there are areas that are manually judged to have fallen off or cracked, but were previously misjudged as areas without problems, adjustments should also be made, and the number of times N that the areas were misjudged as areas without problems is counted. falsenegative And the total number of samples N total , calculate the misjudgment rate If the misjudgment rate is higher than the acceptable threshold T fn , it is necessary to adjust the parameters of the feature extraction link. Taking the edge detection threshold adjustment as an example, let the current edge detection threshold be T edge , the adjusted edge detection threshold is T e ' dge , the following formula can be used: T e ' dge =T edge +γ×(P falsenegative -T fn ), where γ is an adjustment coefficient used to control the amplitude of threshold adjustment. The same idea can be applied to the adjustment of other related parameters.

7. A housing safety monitoring platform based on drone detection according to claim 2, characterized in that: The combined algorithm includes: establishing a two-dimensional rectangular coordinate system on the image plane, taking a corner of the image as the origin, horizontally to the right as the positive direction of the x-axis, and vertically downward as the positive direction of the y-axis, and determining the coordinates (x i ,y i ); Through the system calibration performed in advance, the conversion relationship between the radar scanning space and the image plane coordinate system is established. By placing multiple reference points at known positions, their radar data and position information in the image are obtained at the same time, and then a mathematical model is used to fit the conversion relationship between the two; assuming that the conversion relationship from image coordinates to radar scanning space coordinates obtained after calibration is: Where (x r ,y r ,z r ) are the coordinates of the fixed smear mark point in the radar scanning space, and a1, b1, c1, a2, b2, c2, a3, b3, c3 are the conversion coefficients obtained through calibration.

8. A housing safety monitoring platform based on drone detection according to claim 7, characterized in that: Based on the preprocessed radar data, an algorithm is used to identify the hollow areas in the wall. A reasonable reflection wave intensity threshold T is set. When the intensity of the radar echo signal in a certain area is lower than T, the area is considered to be a hollow area. The boundary extraction and shape fitting of the identified hollow area are performed to obtain the geometric shape description of the hollow area. A polygon can be used to approximate the hollow area, and the coordinate values ​​of its vertex coordinates in the radar scanning space are recorded.

9. A housing safety monitoring platform based on drone detection according to claim 8, characterized in that: Assume that the coordinates of a vertex in the hollow drum area are (x g ,y g ,z g ), the coordinates of the fixed smear mark point in the radar scanning space are (x r ,y r ,z r ), then the distance d from the fixed mark point to the vertex of the hollow area g It can be calculated by the distance formula: Monitoring and tracking in multiple detections g of changes.

Images