Construction site target identification method and system based on neural network

Abstract

The invention discloses a construction site target recognition method and system based on a neural network, and relates to the technical field of image recognition, two infrared cameras are arranged in a preset monitoring area, the shooting angles of the infrared cameras cover the monitoring area, a detection image is shot, and feature points are selected; coordinates of the feature points in the detection image are calibrated in the detection image, and world coordinates are calculated with one feature point as an original point; reconstructing the detection image and extracting the shape of the detection image to obtain a shape image; inputting the shape image into a shape recognition model to obtain a first confidence coefficient; inputting the construction site target brightness into the brightness identification model to obtain a second confidence coefficient; and determining the type of the construction site target based on the first and second confidence coefficients, and determining the three-dimensional coordinates of the construction site target. According to the invention, specific equipment and dangerous targets in a construction site can be accurately identified and positioned.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and more particularly to a construction site target recognition method and system based on a neural network. Background Art

[0002] Safety management is crucial on construction sites. However, traditional site safety monitoring mainly relies on manual inspections, which is not only inefficient but also prone to omissions and misjudgments. With the continuous development of science and technology, people's demand for intelligent site safety management is growing. At present, although there are some safety monitoring systems based on image recognition technology, these systems often find it difficult to accurately locate and identify specific equipment and hazards in complex site environments. For example, fire extinguishers that must be equipped on construction sites may be difficult to find quickly and accurately due to their unstable placement and being blocked by other objects; and for hazardous sources such as wooden ladders that are not allowed on construction sites, they are often difficult to be discovered and handled in a timely manner due to the complexity and variability of the site environment.

[0003] Therefore, there is an urgent need for a technical means that can accurately locate and identify specific equipment and hazardous sources in complex construction site environments. Summary of the invention

[0004] In view of this, the present invention provides a construction site target recognition method and system based on a neural network, which can accurately identify and locate specific equipment and dangerous targets in the construction site.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] A construction site target recognition method based on a neural network, wherein two infrared cameras are arranged in a preset monitoring area, and the shooting angles of the infrared cameras cover the monitoring area, including:

[0007] The first detection image and the second detection image are acquired by shooting with the two infrared cameras, a preset number of identical feature points are selected in the area where the detection images are located, and the coordinates of the identical feature points in the first detection image are calibrated in the first detection image, recorded as first coordinates; the image coordinates of the identical feature points in the second detection image are calibrated in the second detection image, recorded as second coordinates; one identical feature point is set as the origin, and the world coordinates of the preset number of identical feature points are calculated;

[0008] Performing image preprocessing on each detection image, reconstructing the preprocessed detection image to obtain a reconstructed image, converting the reconstructed image into a grayscale image, performing shape extraction on the grayscale image to obtain a shape image;

[0009] Input the appearance image into a pre-trained appearance recognition model to obtain a first confidence level of the construction site target; input the average brightness of each part of the construction site target into a pre-trained brightness recognition model to obtain a second confidence level of the construction site target;

[0010] Based on the first confidence level and the second confidence level, the type of the construction site target is determined, and an anchor point is created according to the world coordinates of the same feature point, and the coordinates of the anchor point are used as the three-dimensional coordinates of the construction site target; wherein the anchor point is the middle point of the preset number of feature points.

[0011] Preferably, the selecting of a preset number of identical feature points specifically includes obtaining a rectangular frame of the detection image in the picture through a target detection algorithm; selecting five feature points within the rectangular frame according to a preset rule, wherein the preset rule is: taking the upper left corner of the picture as the origin of the coordinate system, the coordinates of the upper left corner of the rectangular frame are (A, B), and the coordinates of the lower right corner of the rectangular frame are (C, D), then the coordinates of the five selected feature points are and

[0012] Preferably, the image preprocessing of each detection image specifically includes performing a multi-layer discrete stationary wavelet transform on the detection image to obtain low-frequency sub-band images of multiple scales and high-frequency sub-band images of multiple scales; performing signal enhancement processing on the high-frequency sub-band images and the low-frequency sub-band images, and performing an inverse stationary wavelet transform based on the high-frequency sub-band images and the low-frequency sub-band images after the signal enhancement processing to obtain a reconstructed image.

[0013] Preferably, the signal enhancement processing of the high-frequency sub-band image specifically includes: determining a first segmentation threshold for signal and noise segmentation; segmenting the high-frequency wavelet coefficients in the high-frequency sub-band image of each scale based on the first segmentation threshold to identify valid points and noise points in the high-frequency sub-band image of each scale; and enhancing the high-frequency wavelet coefficient amplitude of each pixel point using different enhancement coefficients.

[0014] Preferably, performing signal enhancement processing on the low-frequency sub-band image specifically comprises: calculating the average value of the low-frequency wavelet coefficient amplitude of the pixel points in the low-frequency sub-band image of each scale; identifying the background area and the target area in the low-frequency sub-band image of each scale based on the average value;

[0015] The average value of the amplitude of the low-frequency wavelet coefficient corresponding to the background area is used as the second segmentation threshold, and the low-frequency wavelet coefficients in the low-frequency sub-band image of each scale are segmented to identify the valid points and noise points in the low-frequency sub-band image of each scale;

[0016] Based on the absolute values ​​of all low-frequency wavelet coefficient amplitudes, the enhancement coefficients corresponding to the low-frequency sub-band images of each scale are determined, and the low-frequency wavelet coefficient amplitudes of each pixel in the low-frequency sub-band images of each scale are enhanced by the enhancement coefficients.

[0017] Preferably, the method further includes calibrating the coordinates of the same feature points in the detection image, calculating the distances between the preset number of same feature points and the camera according to the coordinates, and judging whether the positioning of the detection image is effective according to the calculated distances, which specifically includes:

[0018] Calculate a mathematical expectation value according to the obtained spacing, the mathematical expectation value being the average value of the distances between the preset number of identical feature points and the camera, and calculate a standard deviation of the obtained spacing and the mathematical expectation value;

[0019] Determine whether the standard deviation exceeds a preset threshold value, so as to determine whether the positioning of the detection image through the preset number of feature points is valid, wherein when the value of the standard deviation is greater than the threshold value, it is determined that the positioning of the detection image through the preset number of same feature points is invalid; when the value of the standard deviation is less than or equal to the threshold value, it is determined that the positioning of the detection image through the preset number of same feature points is valid.

[0020] Preferably, the method further comprises calculating a first mapping relationship according to the first coordinates and the world coordinates of the same feature point, and calculating a second mapping relationship according to the second coordinates and the world coordinates of the same feature point;

[0021] Calibrate the first target plane coordinates of the construction site target, and use the first target plane coordinates and the first mapping relationship to calculate the first restored world coordinates of the construction site target; calibrate the second target plane coordinates of the construction site target, and use the second target plane coordinates and the second mapping relationship to calculate the second restored world coordinates of the construction site target;

[0022] Compare and determine whether the deviation between the first restored world coordinate and the second restored world coordinate belongs to the set interval, if so, output the recognition number, if not, output the correction number, the correction number = recognition number + 1.

[0023] A construction site target recognition system based on a neural network, wherein two infrared cameras are arranged in a preset monitoring area, and the shooting angles of the infrared cameras cover the monitoring area, including:

[0024] A coordinate calculation module, which obtains a first detection image and a second detection image by shooting with the two infrared cameras, selects a preset number of identical feature points in the area where the detection images are located, calibrates the coordinates of the identical feature points in the first detection image in the first detection image, and records them as first coordinates; calibrates the image coordinates of the identical feature points in the second detection image in the second detection image, and records them as second coordinates; sets one identical feature point as the origin, and calculates the world coordinates of the preset number of identical feature points;

[0025] An image extraction module performs image preprocessing on each detection image, reconstructs the detection image after preprocessing to obtain a reconstructed image, converts the reconstructed image into a grayscale image, performs shape extraction on the grayscale image, and obtains a shape image;

[0026] The confidence calculation module inputs the appearance image into a pre-trained appearance recognition model to obtain a first confidence of the construction site target; inputs the average brightness of each part of the construction site target into the pre-trained brightness recognition model to obtain a second confidence of the construction site target;

[0027] The target output module determines the type of the construction site target based on the first confidence level and the second confidence level, and creates an anchor point according to the world coordinates of the same feature point, and uses the coordinates of the anchor point as the three-dimensional coordinates of the construction site target; wherein the anchor point is the middle point of the preset number of feature points.

[0028] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for identifying construction site targets based on a neural network, which combines image appearance and brightness information, calculates confidence by using an appearance recognition model and a brightness recognition model to determine the type of construction site targets, and can identify targets more comprehensively and accurately; by calculating the world coordinates of the same feature points and creating anchor points to determine the three-dimensional coordinates of the construction site targets, it is possible to accurately locate the construction site targets in space, which helps to more accurately grasp the on-site conditions of the construction site; by using multi-layer discrete stationary wavelet transform and targeted signal enhancement processing to pre-process the detection image, it can effectively improve the image quality, highlight the target features, and thus improve the accuracy of subsequent recognition; it has a judgment mechanism for the effectiveness of the detection image positioning, which can promptly discover possible inaccurate positioning problems and perform corresponding processing to ensure the reliability of the entire recognition system; by calculating and restoring the world coordinates and judging their deviations, the recognition results can be further verified and corrected to improve the accuracy and stability of recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0030] Figure 1 A diagram of the method steps provided by the present invention;

[0031] Figure 2 A schematic diagram of the structure provided by the present invention. DETAILED DESCRIPTION

[0032] 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.

[0033] The embodiment of the present invention discloses a construction site target recognition method based on a neural network, such as Figure 1 As shown, two infrared cameras are set in the preset monitoring area, and the shooting angles of the infrared cameras cover the monitoring area, including:

[0034] The first detection image and the second detection image are obtained by shooting with two infrared cameras, a preset number of identical feature points are selected in the area where the detection images are located, and the coordinates of the identical feature points in the first detection image are calibrated in the first detection image, recorded as the first coordinates; the image coordinates of the identical feature points in the second detection image are calibrated in the second detection image, recorded as the second coordinates; a same feature point is set as the origin, and the world coordinates of a preset number of identical feature points are calculated;

[0035] Perform image preprocessing on each detection image, reconstruct the preprocessed detection image to obtain a reconstructed image, convert the reconstructed image into a grayscale image, perform shape extraction on the grayscale image to obtain a shape image;

[0036] Input the appearance image into the pre-trained appearance recognition model to obtain a first confidence level of the construction site target; input the average brightness of each part of the construction site target into the pre-trained brightness recognition model to obtain a second confidence level of the construction site target;

[0037] Based on the first confidence level and the second confidence level, the type of the construction site target is determined, and at the same time, an anchor point is created according to the world coordinates of the same feature point, and the coordinates of the anchor point are used as the three-dimensional coordinates of the construction site target; wherein the anchor point is the middle point of a preset number of feature points.

[0038] Among them, both the brightness recognition model and the appearance recognition model use convolutional neural networks.

[0039] In a specific embodiment, selecting a preset number of identical feature points specifically includes obtaining a rectangular frame of the detection image in the picture through a target detection algorithm; selecting five feature points within the rectangular frame according to a preset rule, wherein the preset rule is: taking the upper left corner of the picture as the origin of the coordinate system, the coordinates of the upper left corner of the rectangular frame are (A, B), and the coordinates of the lower right corner of the rectangular frame are (C, D), then the coordinates of the five selected feature points are and

[0040] In a specific embodiment, image preprocessing of each detection image specifically includes performing a multi-layer discrete stationary wavelet transform on the detection image to obtain low-frequency sub-band images of multiple scales and high-frequency sub-band images of multiple scales; performing signal enhancement processing on the high-frequency sub-band image and the low-frequency sub-band image, and performing an inverse stationary wavelet transform based on the high-frequency sub-band image and the low-frequency sub-band image after the signal enhancement processing to obtain a reconstructed image.

[0041] In a specific embodiment, signal enhancement processing of the high-frequency sub-band image specifically includes: determining a first segmentation threshold for signal and noise segmentation; segmenting the high-frequency wavelet coefficients in the high-frequency sub-band images of each scale based on the first segmentation threshold to identify valid points and noise points in the high-frequency sub-band images of each scale; and enhancing the amplitude of the high-frequency wavelet coefficient of each pixel point using different enhancement coefficients.

[0042] The following formula is used to enhance the high-frequency wavelet coefficient amplitude of each pixel in the high-frequency sub-band image of each scale:

[0043]

[0044] Where W is the first segmentation threshold, α1 and α2 are enhancement coefficients, and α1>1, 0<α2<1, is the high-frequency wavelet coefficient amplitude of the pixel with pixel coordinate (x, y) in the high-frequency subband image of the jth scale in the Uth direction before enhancement processing, It is the amplitude of the high-frequency wavelet coefficient of the pixel with pixel coordinate (x, y) in the high-frequency sub-band image of the j-th scale in the U-th direction after enhancement processing.

[0045] In a specific embodiment, performing signal enhancement processing on the low-frequency sub-band image specifically includes: calculating the average value of the low-frequency wavelet coefficient amplitude of the pixel points in the low-frequency sub-band image of each scale; identifying the background area and the target area in the low-frequency sub-band image of each scale based on the average value;

[0046] The average value of the amplitude of the low-frequency wavelet coefficient corresponding to the background area is used as the second segmentation threshold, and the low-frequency wavelet coefficients in the low-frequency sub-band images of each scale are segmented to identify the valid points and noise points in the low-frequency sub-band images of each scale;

[0047] Based on the absolute values ​​of all low-frequency wavelet coefficient amplitudes, the enhancement coefficients corresponding to the low-frequency sub-band images of each scale are determined, and the low-frequency wavelet coefficient amplitudes of each pixel in the low-frequency sub-band images of each scale are enhanced by the enhancement coefficients.

[0048] Among them, the following formula is used to enhance the low-frequency wavelet coefficient amplitude of each pixel in the low-frequency sub-band image of each scale:

[0049]

[0050] Among them, sign() is the sign function, is the low-frequency wavelet coefficient amplitude of the pixel with pixel coordinate (x, y) in the low-frequency sub-band image of the jth scale before enhancement processing, is the amplitude of the low-frequency wavelet coefficient of the pixel with the pixel coordinate (x, y) in the low-frequency sub-band image of the jth scale after enhancement processing, W j is the second segmentation threshold corresponding to the low-frequency sub-band image of the jth scale, k is an exponential factor,

[0051] In a specific embodiment, the method further includes calibrating the coordinates of the same feature points in the detection image, calculating the distances between a preset number of the same feature points and the camera according to the coordinates, and judging whether the positioning of the detection image is effective according to the calculated distances, which specifically includes:

[0052] Calculate a mathematical expectation value based on the obtained spacing, the mathematical expectation value is the average value of the distance between a preset number of identical feature points and the camera, and calculate the standard deviation of the obtained spacing and the mathematical expectation value;

[0053] Determine whether the standard deviation exceeds a preset threshold value, so as to determine whether the positioning of the detection image by a preset number of feature points is effective. When the value of the standard deviation is greater than the threshold value, it is determined that the positioning of the detection image by a preset number of the same feature points is invalid; when the value of the standard deviation is less than or equal to the threshold value, it is determined that the positioning of the detection image by a preset number of the same feature points is effective.

[0054] In a specific embodiment, the first mapping relationship is calculated according to the first coordinates and the world coordinates of the same feature point, and the second mapping relationship is calculated according to the second coordinates and the world coordinates of the same feature point;

[0055] Calibrate the first target plane coordinates of the construction site target, and use the first target plane coordinates and the first mapping relationship to calculate the first restored world coordinates of the construction site target; calibrate the second target plane coordinates of the construction site target, and use the second target plane coordinates and the second mapping relationship to calculate the second restored world coordinates of the construction site target;

[0056] Compare and determine whether the deviation between the first restored world coordinates and the second restored world coordinates belongs to the set interval. If so, output the recognition number; if not, output the correction number, where the correction number = recognition number + 1.

[0057] The first mapping relationship refers to the mapping projection from one plane to another plane in computer vision. That is, the projection mapping relationship between the plane of the monitoring site and the imaging plane of the first camera. Similarly, the second mapping relationship is also the mapping projection from one plane to another plane in computer vision, that is, the projection mapping relationship between the plane of the monitoring site and the imaging plane of the second camera. Taking the first coordinate as an example, the first coordinate expressed in homogeneous coordinates is m = [u, v, 1] T , the world coordinate expressed in homogeneous coordinates is X = [X, Y, Z, 1] T , the homography relationship between the two is s0m=K[R,T]X, where s0 is the scale factor, K is the intrinsic parameter of the first camera, R is the rotation matrix, T is the translation vector, α, γ, β are constants, and [u0,v0,1] is the homogeneous coordinate of the plane with Z=0.

[0058] make Set the monitoring site plane as the plane with Z=0;

[0059] Then we can get

[0060] Define K[r1 r2 t] as the homography matrix H1, that is, the first mapping relationship; using a similar method, the second mapping relationship H2 can also be calculated.

[0061] The first confidence level P of the construction site target s1 and the second confidence level P of the construction site target s2 The following formula is used for fusion to obtain the final category prediction probability:

[0062] P class =max(P s1 ,P s2 );

[0063] When the probability value is greater than the threshold value of 0.5, the target is considered to be correctly identified.

[0064] A construction site target recognition system based on neural network, such as Figure 2As shown, two infrared cameras are set in the preset monitoring area, and the shooting angles of the infrared cameras cover the monitoring area, including:

[0065] A coordinate calculation module is provided, which obtains a first detection image and a second detection image by shooting with two infrared cameras, selects a preset number of identical feature points in the area where the detection images are located, calibrates the coordinates of the identical feature points in the first detection image in the first detection image, and records them as first coordinates; calibrates the image coordinates of the identical feature points in the second detection image in the second detection image, and records them as second coordinates; sets one identical feature point as the origin, and calculates the world coordinates of a preset number of identical feature points;

[0066] The image extraction module performs image preprocessing on each detection image, reconstructs the preprocessed detection image to obtain a reconstructed image, converts the reconstructed image into a grayscale image, performs shape extraction on the grayscale image, and obtains a shape image;

[0067] The confidence calculation module inputs the appearance image into the pre-trained appearance recognition model to obtain the first confidence of the construction site target; inputs the average brightness of each part of the construction site target into the pre-trained brightness recognition model to obtain the second confidence of the construction site target;

[0068] The target output module determines the type of the construction site target based on the first confidence level and the second confidence level, and creates an anchor point according to the world coordinates of the same feature point, and uses the coordinates of the anchor point as the three-dimensional coordinates of the construction site target; wherein the anchor point is the middle point of a preset number of feature points.

[0069] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0070] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A construction site target recognition method based on a neural network, wherein two infrared cameras are arranged in a preset monitoring area, and the shooting angles of the infrared cameras cover the monitoring area, characterized in that: include: The first detection image and the second detection image are acquired by shooting with the two infrared cameras, a preset number of identical feature points are selected in the area where the detection images are located, and the coordinates of the identical feature points in the first detection image are calibrated in the first detection image, which are recorded as first coordinates; and the image coordinates of the identical feature points in the second detection image are calibrated in the second detection image, which are recorded as second coordinates; Setting a same feature point as the origin, and calculating the world coordinates of the preset number of same feature points; Performing image preprocessing on each detection image, reconstructing the preprocessed detection image to obtain a reconstructed image, converting the reconstructed image into a grayscale image, performing shape extraction on the grayscale image to obtain a shape image; Inputting the appearance image into a pre-trained appearance recognition model to obtain a first confidence level of the construction site target; Inputting the average brightness of each part of the construction site target into a pre-trained brightness recognition model to obtain a second confidence level of the construction site target; Based on the first confidence level and the second confidence level, the type of the construction site target is determined, and an anchor point is created according to the world coordinates of the same feature point, and the coordinates of the anchor point are used as the three-dimensional coordinates of the construction site target; wherein the anchor point is the middle point of the preset number of feature points.

2. A method for identifying construction site targets based on a neural network according to claim 1, characterized in that: The selecting of a preset number of identical feature points specifically includes obtaining a rectangular frame of the detection image in the picture through a target detection algorithm; selecting five feature points within the rectangular frame according to a preset rule, wherein the preset rule is: taking the upper left corner of the picture as the origin of the coordinate system, the coordinates of the upper left corner of the rectangular frame are (A, B), and the coordinates of the lower right corner of the rectangular frame are (C, D), then the coordinates of the five selected feature points are and 3. A method for identifying construction site targets based on a neural network according to claim 1, characterized in that: The image preprocessing of each detection image specifically includes performing a multi-layer discrete stationary wavelet transform on the detection image to obtain low-frequency sub-band images of multiple scales and high-frequency sub-band images of multiple scales; performing signal enhancement processing on the high-frequency sub-band image and the low-frequency sub-band image, and performing an inverse stationary wavelet transform based on the high-frequency sub-band image and the low-frequency sub-band image after the signal enhancement processing to obtain a reconstructed image.

4. A method for identifying construction site targets based on a neural network according to claim 3, characterized in that: The signal enhancement processing of the high-frequency sub-band image specifically includes: determining a first segmentation threshold for signal and noise segmentation; segmenting the high-frequency wavelet coefficients in the high-frequency sub-band image of each scale based on the first segmentation threshold to identify valid points and noise points in the high-frequency sub-band image of each scale; and enhancing the high-frequency wavelet coefficient amplitude of each pixel point using different enhancement coefficients.

5. A method for identifying construction site targets based on a neural network according to claim 3, characterized in that: The signal enhancement processing of the low-frequency sub-band image specifically includes: calculating the average value of the low-frequency wavelet coefficient amplitude of the pixel points in the low-frequency sub-band image of each scale; identifying the background area and the target area in the low-frequency sub-band image of each scale based on the average value; The average value of the amplitude of the low-frequency wavelet coefficient corresponding to the background area is used as the second segmentation threshold, and the low-frequency wavelet coefficients in the low-frequency sub-band image of each scale are segmented to identify the valid points and noise points in the low-frequency sub-band image of each scale; Based on the absolute values ​​of all low-frequency wavelet coefficient amplitudes, the enhancement coefficients corresponding to the low-frequency sub-band images of each scale are determined, and the low-frequency wavelet coefficient amplitudes of each pixel in the low-frequency sub-band images of each scale are enhanced by the enhancement coefficients.

6. A method for identifying construction site targets based on a neural network according to claim 1, characterized in that: The method further includes calibrating the coordinates of the same feature points in the detection image, calculating the distances between the preset number of the same feature points and the camera according to the coordinates, and judging whether the positioning of the detection image is effective according to the calculated distances, which specifically includes: Calculate a mathematical expectation value according to the obtained spacing, the mathematical expectation value being the average value of the distances between the preset number of identical feature points and the camera, and calculate a standard deviation of the obtained spacing and the mathematical expectation value; Determine whether the standard deviation exceeds a preset threshold value, so as to determine whether the positioning of the detection image through the preset number of feature points is valid, wherein when the value of the standard deviation is greater than the threshold value, it is determined that the positioning of the detection image through the preset number of same feature points is invalid; when the value of the standard deviation is less than or equal to the threshold value, it is determined that the positioning of the detection image through the preset number of same feature points is valid.

7. A method for identifying construction site targets based on a neural network according to claim 1, characterized in that: The method further comprises calculating a first mapping relationship according to the first coordinates and the world coordinates of the same feature point, and calculating a second mapping relationship according to the second coordinates and the world coordinates of the same feature point; Calibrate the first target plane coordinates of the construction site target, and use the first target plane coordinates and the first mapping relationship to calculate the first restored world coordinates of the construction site target; calibrate the second target plane coordinates of the construction site target, and use the second target plane coordinates and the second mapping relationship to calculate the second restored world coordinates of the construction site target; Compare and determine whether the deviation between the first restored world coordinate and the second restored world coordinate belongs to the set interval, if so, output the recognition number, if not, output the correction number, the correction number = recognition number + 1.

8. A construction site target recognition system based on a neural network, applying the construction site target recognition method based on a neural network described in any one of claims 1 to 7, setting two infrared cameras in a preset monitoring area, the shooting angles of the infrared cameras covering the monitoring area, characterized in that: include: A coordinate calculation module, which obtains a first detection image and a second detection image by shooting with the two infrared cameras, selects a preset number of identical feature points in the area where the detection images are located, calibrates the coordinates of the identical feature points in the first detection image in the first detection image, and records them as first coordinates; calibrates the image coordinates of the identical feature points in the second detection image in the second detection image, and records them as second coordinates; sets one identical feature point as the origin, and calculates the world coordinates of the preset number of identical feature points; An image extraction module performs image preprocessing on each detection image, reconstructs the detection image after preprocessing to obtain a reconstructed image, converts the reconstructed image into a grayscale image, performs shape extraction on the grayscale image, and obtains a shape image; A confidence calculation module, inputting the appearance image into a pre-trained appearance recognition model to obtain a first confidence of the construction site target; Inputting the average brightness of each part of the construction site target into a pre-trained brightness recognition model to obtain a second confidence level of the construction site target; The target output module determines the type of the construction site target based on the first confidence level and the second confidence level, and creates an anchor point according to the world coordinates of the same feature point, and uses the coordinates of the anchor point as the three-dimensional coordinates of the construction site target; wherein the anchor point is the middle point of the preset number of feature points.

Images