Unmanned aerial vehicle bridge deflection detection method based on improved YOLOv8

Abstract

The invention discloses an unmanned aerial vehicle bridge deflection detection method based on improved YOLOv8, establishes a bridge deflection detection solution based on unmanned aerial vehicle photography, uses an unmanned aerial vehicle to shoot a bridge picture, and accurately measures the bridge deflection and eliminates the influence of unmanned aerial vehicle vibration through a target detection method based on deep learning. Specifically, a measurement mark is designed, an unmanned aerial vehicle is utilized to shoot measurement mark photos under different conditions, a target tracking technology based on PSA-YOLOv8 deep learning is introduced to position a mark position, and a Harris-Hough detection algorithm is utilized to calculate a mark center coordinate. And correcting according to an unmanned aerial vehicle vibration elimination algorithm based on a reference balance compensation method to obtain an accurate coordinate of the mark center, so as to calculate the accurate deflection of the bridge. The method has the characteristics of high precision, high robustness, no contact, low cost and the like, and can be suitable for bridge deflection detection under different complex backgrounds.

Description

Technical Field

[0001] The invention relates to the field of bridge displacement detection, and in particular to an unmanned aerial vehicle bridge deflection detection method based on improved YOLOv8. Background Art

[0002] Ensuring the safety and reliability of bridges is the key to ensuring smooth traffic and the safety of people's lives and property. However, over time, many bridges may face various damage and aging problems, including bridge displacement detection, which requires a lot of bridge maintenance work. Traditional bridge deflection detection methods usually require inspectors to carry equipment to climb or use hanging baskets, which is not only time-consuming but also inefficient. At the same time, traditional deflection detection methods face safety threats caused by high-altitude operations when detecting tall and complex bridge structures. With the continuous development and maturity of computer technology, the application of computer vision technology in bridge deflection detection has gradually become possible.

[0003] At present, the bridge structure displacement measurement system based on computer vision generally uses a camera to acquire digital images, and uses a computer and corresponding image processing algorithms to process the image data, and finally obtains the vibration displacement curve of the bridge structure.

[0004] However, bridges are usually located in outdoor environments and are affected by natural factors such as wind, rain, and snow. These factors will affect the stability and measurement accuracy of fixed cameras, increasing the difficulty of detection. Drones can work in various severe weather conditions, such as strong winds, rain, snow, etc. In addition, drones are not constrained by terrain and can operate in complex terrains such as mountainous areas and plateaus. Drones can also quickly reach various parts of the bridge, greatly shortening the detection time and providing a more efficient way for bridge detection. Therefore, it is particularly critical to invent a high-precision detection method for bridge deflection using drones. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a UAV bridge deflection detection method based on improved YOLOv8, which uses a UAV to take pictures of the bridge, performs target detection through the PSA-YOLOv8 algorithm model, and uses the Harris-Hough detection algorithm to calculate the marker center coordinates. The UAV vibration elimination algorithm based on the reference balance compensation method is used to reduce the influence of the UAV's own vibration on the result, thereby achieving high-precision bridge deflection detection.

[0006] Technical solution: The present invention provides a UAV bridge deflection detection method based on improved YOLOv8, the simulation field map is as follows Figure 1 As shown in the figure, the technical method is as follows Figure 3 As shown, the following steps are included:

[0007] S1: Design a Figure 2 Measurement marks shown.

[0008] S2: Take measurement and marking photos under different conditions, annotate key information, obtain the marked data set, and divide it into training set and test set.

[0009] S3: Build a PSA-YOLOv8 target detection network model. The PSA-YOLOv8 network model uses the YOLOv8 target detection network as the basic network and integrates the Pyramid Split Attention (PSA) attention mechanism module into the backbone network of the YOLOv8 network.

[0010] S4: Train the improved PSA-YOLOv8 object detection model using the labeled training set.

[0011] S5: The improved PSA YOLOv8 target detection network is inferred on the test set: PSA-YOLOv8 recognizes the target and obtains the marked detection box information, and finally obtains accurate detection results.

[0012] S6: Gaussian filtering is used to reduce noise in the image, and then the Harris-Hough detection algorithm is used to calculate the coordinates of the marker center.

[0013] S7: Establish a model for drone vibration elimination based on the reference balance compensation method, and correct the marker center coordinates according to the model, and then calculate the actual precise deflection based on the coordinates.

[0014] In S1, Figure 2 As shown in the figure, the measurement mark shape used is composed of four black rectangular blocks forming a cross-shaped area and a black circle in the center. The marking material is made of PVC plastic plate. The target pattern is sprayed on the surface of the plastic plate by a precision inkjet printer, and then covered with an anti-oxidation transparent protective layer to slow down the fading of the target. The mark consists of a reference mark and a measurement mark: the one set at the two end supports of the bridge is called a reference mark, and the one set at the part to be measured on the bridge is called a measurement mark.

[0015] In S2, the captured pictures take into account the influence of different external complex interference environments, thereby enhancing the application scope and applicability of the model under various conditions.

[0016] In S3, the constructed PSA-YOLOv8 target detection model can extract richer feature information than YOLOv8. Compared with other attention modules, the PSA module is lightweight, simple and efficient, which effectively improves the performance of the model. The PSA module is mainly implemented in four steps, such as Figure 4As shown. First, by implementing the proposed Squeeze and Concat (SPC) module, the multi-scale feature map on the channel is obtained. Secondly, the SEWeight module is used to extract the attention of feature maps of different scales and obtain the attention vector in the channel direction. Thirdly, the recalibrated weights of the multi-scale channels are obtained by recalibrating the channel attention vector using Softmax. Fourthly, the element-wise product operation is applied to the recalibrated weights and the corresponding feature map. Finally, a feature map with richer multi-scale feature information can be obtained as output.

[0017] In S4, the training method of the PSA-YOLOv8 network model is: dividing the data set into a training set and a validation set, setting the training coefficient and the number of images input for each training, and saving the trained weights after the training is completed.

[0018] In S6, the Harris-Hough detection algorithm is used to calculate the center coordinates of the marker. Specifically, the Harris corner detection algorithm is used to detect and match corners, output the corner coordinates, and use the Hough transform to fit the center of the circle to determine the center coordinates of the target circle. The center coordinates and the corner coordinates are averaged to obtain the precise center coordinates, thereby calculating the actual precise deflection of the bridge. Harris corner detection is an algorithm for detecting corners in an image in computer vision. The specific formula is as follows:

[0019] E(u,v)=∑ x,y ω(x,y)[I(x+u,y+v)-I(x,y)] 2

[0020] Among them, u and v represent the offset in the vertical and horizontal directions respectively, ω(x,y) is the center of the pixel frame; I(x+u,y+v) is the grayscale value of the coordinates plus the offset (x+u,y+v) after the pixel frame is shifted, and I(x,y) is the grayscale value of the coordinate position (x,y) before the shift; E represents the sum of the squares of the differences between the corresponding positions of each pixel selected before the shift and each pixel selected after the shift.

[0021] Then, the iterative interpolation method is used to obtain the sub-pixel corner coordinates, and the selected corner points are approximated by a quadratic polynomial, which is defined as the response function R(x,y):

[0022] ax 2 +by 2 +cxy+dx+ey=R(x,y)

[0023] The above method has located the coordinates of the corner point (x, y). Subsequently, 9 pixels in the surrounding neighborhood are found with the pixel point as the center, and a set of equations is established and solved by the least squares method. After that, R(x, y) is derived to obtain the coordinates of the corner points of each rectangle, so that a target center coordinate α(x1, y1) can be determined. For the center circle, the Hough gradient method in the Hough transform is used to detect the center coordinates of the circle. Hough gradient circle detection locates the center of the circle based on the gradient information of the image, and then calculates the radius of the circle. The advantage of this method is that it reduces the calculation dimension, thereby reducing the amount of calculation and detection time. The Hough gradient circle detection process is as follows Figure 7 As shown in a, b, c, and d, the gradient direction of all pixels is solved by the Sobel operator, and the normal through the point is drawn from the gradient direction of the pixel. If the pixel is on the circle, the normal through the point is the tangent of the circle, and the gradient direction of the point points to the center of the circle. As shown in Figure a, in the XOY plane coordinate system, the normals of the detected points A, B, C, D, and E are drawn. As shown in Figure b, a straight line is drawn along the gradient direction of the above 5 points. The point where the perpendicular lines intersect is the potential center of the circle. In this way, a straight line is drawn along the gradient direction of all pixels. The point with the most intersections is found, which is the potential center of the circle. Point P is the intersection of 4 perpendicular lines, and points P1, P2, and P3 are all the intersections of 2 perpendicular lines. Therefore, the number of votes for the center of the circle P is 4, and the number of votes for the centers of the circle P1, P2, and P3 are all 2, so point P is the final center of the circle. As shown in Figure c, suppose the distances from points A, B, and C to the intersection point P are all r1, and the distance from point D to the intersection point P is r2. In the circle with P as the center, the number of votes for the radius r1 is 3, and the number of votes for the radius r2 is 1. This method is used to find the radius with the most votes. With this center and radius, a Hough gradient circle can be obtained, and the center of the circle is set as β(x2, y2). Finally, the midpoint coordinate formula is applied to the two extracted coordinate points:

[0024]

[0025] Get the target center coordinates γ′(x′0, y′0). Through the above method, we will get the coordinates of two reference marks γ′1(x′1, y′1), γ′2(x′2, y′2) and the coordinates of multiple positions to be measured.

[0026] In S7, a model for eliminating drone vibration based on the reference balance compensation method is established, and the coordinates are calibrated using plane isomorphism transformation, which effectively reduces the impact of the drone's own vibration on the measurement results. Due to the vibration of the drone and the different shooting times, the positions of the captured images at two moments will also be different, and this difference can be described by marking the center coordinates. The center plane coordinates of the measurement mark before displacement are expressed as:

[0027] γ0(x0,y0)

[0028] In addition, the center plane coordinates of the measured marker after displacement are expressed as:

[0029] γ′ t (x′ t ,y′ t )

[0030] We describe the center point of the above marker as the spatial coordinates [x0, y0, 1] T and [x′ t ,y′ t ,1] T Next, we use the plane isomorphism transformation matrix to compensate for the camera motion, which can be expressed as:

[0031]

[0032] The parameter θ is the rotation angle, while dx and dy represent the x-axis and y-axis displacements, respectively. The additional term s represents the isotropic scale factor. In the above equation, there are four unknown parameters. The coordinates of the two reference markers γ′1(x′1, y′1) and γ′2(x′2, y′2) can provide two sets of non-collinear points, thus satisfying the requirements for solving the system of equations. By solving the system of equations, we obtain a 3×3 transformation matrix. To simplify the notation, we denote the 3×3 transformation matrix as M. The compensated measurement marker coordinates can be calculated by the following formula:

[0033] [x t ,y t ,1] T =M[x′ t ,y′ t ,1]

[0034] Through the above process, each subsequent marking image taken by the drone can be aligned with the first one, so as to accurately determine the center coordinates γ of the measurement mark t (x t ,y t ), and the relative displacement value is calculated with high precision by the following formula:

[0035] d x =x0-x t , d y =y0-y t

[0036] Since the reference marker coordinates reflect the movement of the drone, and the coordinates of the reference marker after displacement are γ″1(x″1,y″1), γ″2(x″2,y″2), the reference system displacement d′ can be calculated based on the reference coordinates. x ,d′ y :

[0037]

[0038] Finally, the relative displacement value is subtracted from the reference system displacement to obtain the accurate absolute displacement value:

[0039] D′ x =d x -d′ x , D′ y =d y -d′ y

[0040] The above displacement values ​​are calculated based on pixel displacement values. To calculate the actual displacement value of the bridge, it is necessary to adjust the ratio. The actual distance between the corner points of the measurement mark in the horizontal and vertical directions is known to be L. x ,L y , the horizontal distance of the corner pixels calculated on the image is recorded as Δ x , similarly, the average vertical distance is recorded as Δ y , then the horizontal scale conversion factor K x and the vertical scale conversion factor K y They are:

[0041]

[0042] The actual displacement of the bridge is:

[0043] D x =K x ·D′ x , D y =K y ·D′ y

[0044] The beneficial effects of the present invention are:

[0045] The present invention designs a UAV bridge deflection detection method based on improved YOLOv8, uses a UAV to take measurement mark photos under different conditions, performs target detection through the PSA-YOLOv8 algorithm model, further uses the Harris-Hough detection algorithm to calculate the mark center coordinates, and uses the UAV vibration elimination algorithm based on the reference balance compensation method to reduce the influence of the UAV's own vibration on the measurement results, so as to accurately calculate the bridge deflection according to the corrected coordinates. The UAV bridge deflection detection method based on improved YOLOv8 of the present invention can realize millimeter-level detection of bridge deflection under complex terrain background conditions, and requires less equipment and operators. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a schematic diagram of the simulated site designed by the present invention.

[0047] Figure 2 This is a schematic diagram of a measurement mark designed by the present invention.

[0048] Figure 3 It is the technical roadmap designed by the present invention

[0049] Figure 4 Schematic diagram of the PSA module proposed by the present invention

[0050] Figure 5 It is a schematic diagram of the SPC module proposed by the present invention

[0051] Figure 6 This is the network architecture diagram of the bottleneck part in the improved PSA-YOLOv8 proposed by the present invention

[0052] Figure 7 This is the Hough gradient circle detection flow chart proposed by the present invention DETAILED DESCRIPTION

[0053] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0054] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention is further described in detail below.

[0055] The present invention provides a UAV bridge deflection detection method based on improved YOLOv8, comprising the following steps:

[0056] S1: Design a measurement mark. The target is designed mainly based on the feature point information. The designed target is based on the traditional black and white grid target and the black circle. The marking material is made of PVC plastic plate. The target pattern is sprayed on the surface of the plastic plate by a precision inkjet printer, and then covered with an anti-oxidation transparent protective layer to slow down the fading of the target.

[0057] S2: Take measured and marked photos under different conditions, annotate the key information, and obtain a labeled data set. First, use a professional drone to fly under different complex background conditions to capture different visual environments. These pictures under different conditions provide rich visual information for the data set, which is extremely important for training and testing target tracking systems. Then use the open source graphic image annotation tool Labelimg for data annotation. This step uses Labelimg to create bounding boxes and polygon annotations, and it is written in Python and uses Qt as its graphical interface, so it can effectively extract and generate labeled data in YOLO format.

[0058] S3: Construct a PSA-YOLOv8 target detection network model. The PSA YOLOv8 network model uses the YOLOv8 target detection network as the basic network, and integrates the Pyramid Split Attention (PSA) attention mechanism module into the backbone network of the YOLOv8 network. Pyramid Split Attention (PSA) is a module based on the attention mechanism, which can be used in tasks such as image classification and target detection. The PSA module forms a pyramid-shaped feature map by splicing the convolution results of convolution kernels of different sizes, and then applies the attention mechanism on this feature map to extract richer feature information. Compared with other attention modules, the PSA module is lightweight, simple and efficient, thereby improving the performance of the model. The basic operator for implementing multi-scale feature extraction in the proposed PSA is SPC, such as Figure 5 The spatial information of the input feature map is extracted in a multi-branch manner. The input channel dimension of each branch is C. For the input feature map X, the feature map X is first divided into S parts and [X0, X1, …, X S-1 ] to indicate that the number of channels in each segmented part is:

[0059]

[0060] At the same time, the feature map after each segmentation:

[0061] X i ∈R C′×H×W ,i=0,1…,S-1

[0062] R C′×H×W Represents feature maps with different scales. For each divided channel feature map, we apply multi-scale convolution to extract the spatial information of feature maps of different scales. Here we mainly use multi-scale convolution kernel group convolution. The role of group convolution is mainly to reduce the number of parameters. At the same time, we also explored a strategy to adaptively select the group size according to the size of the convolution kernel. The relationship between the group size and the kernel size is:

[0063]

[0064] Therefore, the specific calculation method of the multi-scale feature extraction process is as follows:

[0065] F i =Conv(K i ×K i ,G i )(X i ),i=0,1,2,…,S-1

[0066] Where Conv represents the vector convolution operation, where:

[0067] K i =2×(i+1)+1, F i ∈R C×H×W

[0068] Finally, we can get the feature map after multi-scale fusion:

[0069] F=Cat([F0,F1,…,F S-1 ])

[0070] Here F∈R C×H×W , Cat represents the construction of a multidimensional array. After extracting the multi-scale feature map, we need to perform feature mapping of different scales F i Extract channel attention weights. The specific calculation is as follows:

[0071] Z i =SEWeight(F i ),i=0,1,2,…,S-1

[0072] Where SEWeight represents the vector of attention weights, and Z i ∈R C′×1×1 Then, the entire multi-scale channel attention weight vector is obtained by concatenation:

[0073]

[0074] In the formula Represents the concat operator. In order to establish long-term channel attention dependency and realize information interaction between multi-scale channel attention, we further use Softmax to recalibrate the weights of channel attention information:

[0075]

[0076] Among them, Softmax is used to obtain the recalibration of multi-scale channels, including all position information in space and the attention weights att in the channel i Then, we will correspond to the scale feature map F i Multiply the weighted attention vector channel-wise, that is:

[0077] Y i =F i ⊙att i ,i=1,2,3,…,S-1

[0078] Where att represents the multi-scale channel weight after attention interaction, and ⊙ represents the channel-wise multiplication operation. Finally, we concatenate the feature maps obtained after multi-scale channel attention weighting, and finally output a feature map with richer multi-scale information. The specific calculation is as follows:

[0079] Out=Cat([Y0,Y1,…,Y S-1 ])

[0080] The modified network architecture is shown in the figure below: Figure 6 As shown in the figure, the 3x3 convolution in the bottleneck part of ResNet is mainly replaced by the PSA module.

[0081] S4: Train the improved PSA YOLOv8 target detection model using the labeled training set. The training method of the PSA-YOLOv8 network model is as follows: divide the data set into a training set and a validation set, set the training coefficient and the number of images input for each training, and save the trained weights after the training is completed.

[0082] S5: The improved PSA YOLOv8 target detection network is inferred on the test set: PSA YOLOv8 recognizes the target and obtains the marked detection box information, and finally obtains accurate detection results.

[0083] S6: Use Gaussian filtering technology to reduce noise on the image, and then use the Harris-Hough detection algorithm to calculate the center coordinates of the marker. Specifically, use the Harris corner detection algorithm to detect and match corners, output the corner coordinates, and use the Hough transform to fit the center of the circle to determine the center coordinates of the target circle. The center coordinates and the corner coordinates are averaged to obtain the precise center coordinates, thereby calculating the actual precise deflection of the bridge. Harris corner detection is an algorithm for detecting corners in an image in computer vision. The Harris corner detection principle is transformed into a formula:

[0084] E(u,v)=∑ x,y ω(x,y)[I(x+u,y+v)-I(x,y)] 2

[0085] Among them, u and v represent the offset in the vertical and horizontal directions respectively, ω(x,y) is the center of the pixel frame; I(x+u,y+v) is the grayscale value of the coordinates plus the offset (x+u,y+v) after the pixel frame is shifted, and I(x,y) is the grayscale value of the coordinate position (x,y) before the shift; E represents the sum of the squares of the differences between the corresponding positions of each pixel selected before the shift and each pixel selected after the shift.

[0086] Then, the iterative interpolation method is used to obtain the sub-pixel corner coordinates, and the selected corner points are approximated by a quadratic polynomial, which is defined as the response function R(x,y):

[0087] ax 2 +by 2 +cxy+dx+ey=R(x,y)

[0088] The above method has located the coordinates of the corner point (x, y). Subsequently, 9 pixel points in the surrounding neighborhood are found with the pixel point as the center, and a set of equations is established and solved by the least squares method. After that, R(x, y) is derived to obtain the coordinates of the corner points of each rectangle, thereby determining the center coordinates of the target. For the center circle, the Hough gradient method in the Hough transform is used to detect the center coordinates of the circle. Hough gradient circle detection locates the center of the circle based on the gradient information of the image, and then calculates the radius of the circle. The advantage of this method is that it reduces the calculation dimension, thereby reducing the amount of calculation and detection time. The Hough gradient circle detection process is as follows: Figure 7 As shown in a, b, c, and d, the gradient direction of all pixels is solved by the Sobel operator, and the normal through the point is drawn from the gradient direction of the pixel. If the pixel is on the circle, the normal through the point is the tangent of the circle, and the gradient direction of the point points to the center of the circle. As shown in Figure a, in the XOY plane coordinate system, the normals of the detected points A, B, C, D, and E are drawn. As shown in Figure b, a straight line is drawn along the gradient direction of the above 5 points. The point where the perpendicular lines intersect is the potential center of the circle. In this way, a straight line is drawn along the gradient direction of all pixels. The point with the most intersections is found, which is the potential center of the circle. Point P is the intersection of 4 perpendicular lines, and points P1, P2, and P3 are all the intersections of 2 perpendicular lines. Therefore, the number of votes for the center of the circle P is 4, and the number of votes for the centers of the circle P1, P2, and P3 are all 2, so point P is the final center of the circle. As shown in Figure c, suppose the distances from points A, B, and C to the intersection point P are all r1, and the distance from point D to the intersection point P is r2. In the circle with P as the center, the number of votes for the radius r1 is 3, and the number of votes for the radius r2 is 1. This method is used to find the radius with the most votes. With this center and radius, a Hough gradient circle can be obtained, and the center of the circle is set as β(x2, y2). Finally, the midpoint coordinate formula is applied to the two extracted coordinate points:

[0089]

[0090] The target center coordinates γ′(x′0, y′0) can be obtained. Through the above method, we will get the coordinates of the two reference marks γ′1(x′1, y′1), γ′2(x′2, y′2) and the coordinates of the position to be measured γ′(x′0, y′0).

[0091] S7: Establish a model for drone vibration elimination based on the reference balance compensation method, and correct the marker center coordinates according to the model, and then calculate the actual precise deflection based on the coordinates. The model for drone vibration elimination based on the reference balance compensation method uses plane isomorphism transformation to calibrate the coordinates, which effectively reduces the impact of the drone's own vibration on the measurement results. Due to the vibration of the drone and the different shooting times, the positions of the captured images at two moments will also be different, and this difference can be described by the marker center coordinates. The center plane coordinates of the measurement marker before displacement are expressed as:

[0092] γ0(x0,y0)

[0093] In addition, the center plane coordinates of the measured marker after displacement are expressed as:

[0094] γ′ t (x′ t ,y′ t )

[0095] We describe the center point of the above marker as the spatial coordinates [x0, y0, 1] T and [x′ t ,y′ t ,1] T Next, we use the plane isomorphism transformation matrix to compensate for the camera motion, which can be expressed as:

[0096]

[0097] The parameter θ is the rotation angle, while dx and dy represent the x-axis and y-axis displacements, respectively. The additional term s represents the isotropic scale factor. In the above equation, there are four unknown parameters. The coordinates of the two reference markers γ′1(x′1, y′1) and γ′2(x′2, y′2) can provide two sets of non-collinear points, thus satisfying the requirements for solving the system of equations. By solving the system of equations, we obtain a 3×3 transformation matrix. To simplify the notation, we denote the 3×3 transformation matrix as M. The compensated measurement marker coordinates can be calculated by the following formula:

[0098] [x t ,y t ,1] T =M[x′ t ,y′ t ,1]

[0099] Through the above process, each subsequent marking image taken by the drone can be aligned with the first one, so as to accurately determine the center coordinates γ of the measurement mark t (x t ,y t ), and the relative displacement value is calculated with high precision by the following formula:

[0100] d x =x0-x t , d y =y0-y t

[0101] Since the reference marker coordinates reflect the movement of the drone, and the coordinates of the reference marker after displacement are γ″1(x″1,y″1), γ″2(x″2,y″2), the reference system displacement d′ can be calculated based on the reference coordinates. x ,d′ y :

[0102]

[0103] Finally, the relative displacement value is subtracted from the reference system displacement to obtain the accurate absolute displacement value:

[0104] D′ x =d x -d′ x , D′ y =d y -d′ y

[0105] The above displacement values ​​are calculated based on pixel displacement values. To calculate the actual displacement value of the bridge, it is necessary to adjust the ratio. The actual distance between the corner points of the measurement mark in the horizontal and vertical directions is known to be L. x ,L y , the horizontal distance of the corner pixels calculated on the image is recorded as Δ x , similarly, the average vertical distance is recorded as Δ y , then the horizontal scale conversion factor K x and the vertical scale conversion factor K y They are:

[0106]

[0107] The actual displacement of the bridge is:

[0108] D x =K x ·D′ x , D y =K y ·D′ y

[0109] What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A UAV bridge deflection detection method based on improved YOLOv8, characterized in that: The method comprises the following steps: S1: Design a measurement mark; S2: Use drones to take measurement and marking photos under different conditions, annotate key information, obtain a labeled data set, and divide it into a training set and a test set; S3: construct a PSA-YOLOv8 target detection network model, wherein the PSA-YOLOv8 network model uses the YOLOv8 target detection network as the basic network and integrates a Pyramid Split Attention (PSA) attention mechanism module into the backbone network of the YOLOv8 network; S4: Train the improved PSA-YOLOv8 object detection model using the labeled training set; S5: The improved PSA-YOLOv8 target detection network is inferred on the test set: PSA-YOLOv8 identifies the target and obtains the marked detection box information, and finally obtains accurate detection results; S6: Use Gaussian filtering technique to reduce noise on the image, and then use Harris-Hough detection algorithm to calculate the center coordinates of the marker; S7: Establish a model for drone vibration elimination based on the reference balance compensation method, and correct the marker center coordinates according to the model, and then calculate the actual precise deflection based on the coordinates.

2. According to the method for detecting bridge deflection by unmanned aerial vehicle based on improved YOLOv8 of claim 1, it is characterized in that: This method introduces the target tracking technology based on PSA-YOLOv8 deep learning to locate the marker position, uses the Harris-Hough detection algorithm to calculate the marker center coordinates, and corrects the precise coordinates of the marker center according to an algorithm for drone vibration elimination based on the reference balance compensation method, thereby calculating the precise deflection of the bridge.

3. The unmanned aerial vehicle bridge deflection detection method based on improved YOLOv8 according to claim 1 is characterized in that: In S1, the measurement mark shape used is composed of four black rectangular blocks forming a cross-shaped area and a black circle in the center.

4. The method for detecting bridge deflection by unmanned aerial vehicle based on improved YOLOv8 according to claim 1, characterized in that: In S2, the collected labeled data set takes into account the influence of different external complex interference environments, and expands the application scope of the model under various conditions.

5. The method for detecting bridge deflection using an unmanned aerial vehicle based on improved YOLOv8 according to claim 1, characterized in that: In S3, the constructed PSA-YOLOv8 target detection model is an improved model based on the YOLOv8 model, which effectively improves the performance of the model.

6. The method for detecting bridge deflection using an unmanned aerial vehicle based on improved YOLOv8 according to claim 1, characterized in that: In S4, the training method of the PSA-YOLOv8 network model is: dividing the data set into a training set and a validation set, setting the training coefficient and the number of images input for each training, and saving the trained weights after the training is completed.

7. The method for detecting bridge deflection by unmanned aerial vehicle based on improved YOLOv8 according to claim 1, characterized in that: In S6, the Harris-Hough detection algorithm is used to calculate the marker center coordinates. The Harris corner detection algorithm is used to perform corner detection and matching, output the corner coordinates, and use Hough transform to fit the circle center to determine the target inner circle center coordinates, and the circle center coordinates and corner point coordinates are averaged to obtain the center point coordinates.

8. The method for detecting bridge deflection by unmanned aerial vehicle based on improved YOLOv8 according to claim 1, characterized in that: In S7, a model for eliminating UAV vibration based on the reference balance compensation method is established, the reference coordinates are calibrated using plane isomorphism transformation, the influence of the UAV's own vibration on the measurement results is considered, and finally the actual precise deflection of the bridge is calculated.

Images