A method for identifying, positioning and segmenting road hidden diseases based on three-dimensional ground penetrating radar

Abstract

This invention discloses a method for identifying, locating, and segmenting hidden road defects based on 3D ground-penetrating radar (GPR), belonging to the field of road defect detection technology. This invention addresses the problems of low efficiency, poor accuracy, and inability to automatically reconstruct 3D road defects using traditional methods. By establishing a YOLO-nnUNet 3D GPR defect detection-segmentation joint algorithm, this invention proposes a specific method for rapid location, identification, and 3D segmentation of hidden road defects from 3D GPR data. It realizes a complete defect identification process from 3D GPR data acquisition, 2D defect detection, 3D defect clustering and location, and 3D defect boundary semantic segmentation, improving the efficiency and accuracy of hidden road defect detection. Furthermore, this method can obtain 3D GPR road defect segmentation data and output the 3D reconstruction results of the road defects. This invention can be applied to the detection of hidden road defects.

Description

Technical Field

[0001] This invention belongs to the field of road hidden defects detection technology, specifically relating to a method for identifying, locating and segmenting road hidden defects based on three-dimensional ground penetrating radar. Background Technology

[0002] With the rapid development of transportation infrastructure and the continuous growth of motor vehicle ownership, the frequency of road use and load intensity are constantly increasing. However, hidden road defects (such as voids, cracks, spalling, and cavities) are difficult to detect directly through traditional visual inspection or surface monitoring methods because they are buried under the road surface. Once these defects expand, they will seriously affect the structural safety and durability of the road, and even lead to road collapse, traffic accidents, and other safety hazards. Currently, the main methods for detecting hidden road defects include manual inspection, geological exploration, and radar detection. Among them, three-dimensional ground-penetrating radar detection, as a non-destructive detection technology, can perform high-precision scanning of road structures. However, traditional methods rely on manual experience to analyze radar echo images, making it difficult to achieve automated and intelligent defect detection. Therefore, in order to solve the problems of low efficiency and poor accuracy of traditional methods for detecting hidden road defects, a method that can use intelligent detection methods to replace traditional manual analysis is proposed, which is crucial for promoting the development of road defect detection towards automation, intelligence, and real-time. Moreover, there is currently no technology in traditional methods for automated three-dimensional rapid reconstruction of road defects. Summary of the Invention

[0003] The purpose of this invention is to solve the problems of low efficiency, poor accuracy and inability to automatically reconstruct three-dimensional road defects using traditional methods, and to propose a method for identifying, locating and segmenting road defects based on three-dimensional ground-penetrating radar.

[0004] The technical solution adopted by the present invention to solve the above-mentioned technical problems is: a method for identifying, locating, and segmenting hidden road defects based on three-dimensional ground-penetrating radar, the method specifically including the following steps:

[0005] Step 1: Use a 3D ground-penetrating radar (GPR) device to perform 3D underground space information detection on the road to be inspected, obtain 3D radar signal data, and record the radar signal data corresponding to the coordinate point (i,j,k) as D. i,j,k ;

[0006] Where i∈[0,N], j∈[0,L], k∈[0,M], the three-dimensional ground penetrating radar device moves in a straight line in the horizontal direction. The direction of movement of the three-dimensional ground penetrating radar device is taken as the x-axis direction. N represents the x-axis direction coordinate corresponding to the maximum movement distance of the radar. L represents the y-axis direction coordinate corresponding to the maximum detection width of the radar in the direction perpendicular to the x-axis. M represents the z-axis direction coordinate corresponding to the maximum depth of the radar electromagnetic wave propagation underground.

[0007] By processing the three-dimensional radar signal data, the three-dimensional radar signal data is converted into three-dimensional radar data grayscale blocks;

[0008] Step 2: Slice the grayscale block of the 3D radar data layer by layer along the z-axis to obtain a sequence of 2D grayscale images of the road to be detected. Input the grayscale images in the sequence of 2D grayscale images of the road to be detected one by one into the trained YOLO 2D disease detection model. The YOLO 2D disease detection model outputs the disease detection result for each 2D grayscale image. The disease detection result is the bounding box of the location of the disease in the 2D grayscale image.

[0009] Calculate the center coordinates of the bounding box at the location of each disease, and record the center coordinates of each bounding box. and center coordinates The z-axis coordinates of the corresponding two-dimensional grayscale image;

[0010] Step 3: Calculate the adaptive neighborhood radius ε0 of the DBSCAN algorithm based on the center coordinates of the bounding boxes of the lesions in each 2D grayscale image, and cluster the lesions in each 2D grayscale image based on the adaptive neighborhood radius ε0.

[0011] Patches that are clustered together and spatially connected are merged into a three-dimensional disease candidate region, and the bounding box, center position and size information of each three-dimensional disease candidate region are calculated.

[0012] Step 4: Based on the bounding box, center position, and size information of the three-dimensional defect candidate region, cut out each three-dimensional defect candidate data block from the three-dimensional radar data block of the road to be detected;

[0013] Step 5: Input each 3D disease candidate data block into the trained nnUNet 3D disease segmentation model, perform semantic segmentation on each disease candidate region, and obtain the disease boundary obtained from the semantic segmentation.

[0014] Then, connected component analysis and Gaussian filtering are used to process the disease boundaries extracted by semantic segmentation to obtain the processed disease boundaries.

[0015] Furthermore, the processing of the three-dimensional radar signal data specifically involves:

[0016] The three-dimensional radar signal data is subjected to dynamic compression, data normalization, and grayscale mapping in sequence.

[0017] Furthermore, the process involves processing the three-dimensional radar signal data to convert it into three-dimensional radar data grayscale blocks; the specific process is as follows:

[0018] Step 11: Dynamically compress the acquired 3D radar signal data using logarithmic transformation.

[0019] P i,j,k =log(1+αD) i,j,k )

[0020] Among them, P i,j,k This represents the radar signal data corresponding to coordinate point (i,j,k) after dynamic compression;

[0021] α represents the signal amplification factor;

[0022] Step 1 & 2: Calculate the mean value μ of the radar signal data corresponding to each coordinate point after dynamic compression.

[0023]

[0024] Calculate the standard deviation σ of the radar signal data corresponding to each coordinate point after dynamic compression:

[0025]

[0026] The radar signal data corresponding to each coordinate point are normalized according to the mean μ and standard deviation σ:

[0027]

[0028] Among them, M i,j,k This represents the normalized data corresponding to the coordinate point (i,j,k);

[0029] Step 13: Perform grayscale value mapping on the normalized data corresponding to each coordinate point, and use the grayscale value mapping result of coordinate point (i,j,k) as the grayscale value of voxel (i,j,k):

[0030]

[0031] Wherein, min(M) represents the minimum value in the normalized data corresponding to each coordinate point;

[0032] max(M) represents the maximum value in the normalized data corresponding to each coordinate point;

[0033] H i,j,kThis represents the grayscale value of voxel (i,j,k).

[0034] Furthermore, the center coordinates of the bounding box at the location of the disease are:

[0035]

[0036] in, The minimum x-coordinate of the pixel within the bounding box of the location of the a-th disease;

[0037] The maximum x-coordinate of the pixel within the bounding box representing the location of the a-th disease;

[0038] The minimum ordinate of the pixel within the bounding box representing the location of the a-th disease;

[0039] The maximum ordinate of the pixels within the bounding box representing the location of the a-th disease;

[0040] This represents the center coordinates of the bounding box where the a-th disease is located.

[0041] Furthermore, the method for calculating the adaptive neighborhood radius ε0 is as follows:

[0042] The bounding box center coordinates of the location of the a-th lesion in the two-dimensional grayscale image sequence Calculate the center coordinates of the bounding boxes of other lesions in the two-dimensional grayscale image sequence. The distance is calculated by taking the distance from the given distance and then sorting the calculated distances in ascending order, resulting in the distance vector d. (k) for:

[0043]

[0044] in, Represents the distance sequence d (k) The 1st, 2nd, ..., N'th distance in the sequence;

[0045] N' represents the total number of diseases other than the a-th disease in the two-dimensional grayscale image sequence;

[0046] Based on the distance vector d (k) Get N' pairs of points (x n ,y n ), n=1,2,…,N':

[0047] x n =n,

[0048] Then based on N' pairs of points (x n,y n Obtain the cumulative sum function curve, and then obtain the adaptive neighborhood radius based on the cumulative sum function curve:

[0049]

[0050] Where (x'(n),y'(n)) represents a pair of points (x... n ,y n The first derivative of the cumulative sum function curve, (x)(n), y)(n) represents the pair of points (x...). n ,y n The value of the second derivative on the cumulative sum function curve.

[0051] Furthermore, the specific process of the connected component analysis is as follows:

[0052] Calculate the number of voxels in each connected disease region, and denote the number of voxels in the i'-th connected disease region as V. i’ Then, diseased areas with voxel counts less than a threshold are removed to obtain the remaining diseased areas.

[0053] Furthermore, the specific process of the Gaussian filtering method is as follows:

[0054] Gaussian filtering was used to smooth the boundary voxels of the remaining diseased areas.

[0055]

[0056] Where (x,y,z) represents the boundary voxels of the remaining diseased area;

[0057] H disease (x+p,y+q,z+r) represents the gray value of voxel (x+p,y+q,z+r) before smoothing.

[0058] This represents the Gaussian function value at position (p, q, r).

[0059] H smooth (x,y,z) represents the gray value of the boundary voxel (x,y,z) in the remaining diseased area after smoothing.

[0060] p, q, r represent the values ​​of the Gaussian function. The offsets in the distribution are p, q, r ∈ [-3σ0, 3σ0], where σ0 represents the standard deviation of the Gaussian distribution.

[0061] Furthermore, the Gaussian function value at position (p,q,r) is:

[0062]

[0063] The beneficial effects of this invention are:

[0064] This invention proposes a specific method for the rapid localization, identification, and 3D segmentation of hidden road defects in 3D ground-penetrating radar (GPR) data by establishing a joint YOLO-nnUNet algorithm for 3D GPR defect detection and segmentation. In the task of locating, identifying, and 3D segmenting hidden road defects, considering the small number of defects, their significant impact, and the high efficiency of 2D defect detection, the workflow of the road defect segmentation task is optimized. This achieves a complete defect identification process from 3D GPR data acquisition, 2D defect detection, 3D defect clustering and localization, and 3D defect boundary semantic segmentation, improving the efficiency and accuracy of hidden road defect detection. Furthermore, this method can obtain 3D GPR road defect segmentation data and output the 3D reconstruction results of the road defects.

[0065] This invention provides strong technical support for the automated detection and intelligent assessment of hidden road defects by improving the effective data screening and three-dimensional extraction methods, and provides a data source for road management and maintenance. Attached Figure Description

[0066] Figure 1 This is a flowchart of a method for identifying, locating, and segmenting hidden road defects based on three-dimensional ground-penetrating radar according to the present invention;

[0067] Figure 2 This is a schematic diagram of a two-dimensional annotated image;

[0068] Figure 3 This is a schematic diagram of the 3D voxel-level annotation results;

[0069] Figure 4 It is a slice-like image of the three-dimensional reconstruction result of the disease;

[0070] Figure 5 This is a 3D reconstruction rendering of the disease. Detailed Implementation

[0071] Specific Implementation Method 1: The method for identifying, locating, and segmenting hidden road defects based on three-dimensional ground-penetrating radar described in this implementation method specifically includes the following steps:

[0072] Step 1: Use a 3D ground-penetrating radar (GPR) device to perform 3D underground space information detection on the road to be inspected, obtain 3D radar signal data, and record the radar signal data corresponding to the coordinate point (i,j,k) as D. i,j,x ;

[0073] Where i∈[0,N], j∈[0,L], k∈[0,M], the 3D ground-penetrating radar device moves in a straight line in the horizontal direction. The direction of movement of the 3D ground-penetrating radar device is taken as the x-axis direction. N represents the x-axis coordinate corresponding to the maximum moving distance of the radar, L represents the y-axis coordinate corresponding to the maximum detection width of the radar in the direction perpendicular to the x-axis, and M represents the z-axis coordinate corresponding to the maximum depth of radar electromagnetic wave propagation underground. D i,j,k The radar echo data corresponding to the coordinate point (i,j,k) is used to represent the radar echo data. It should be noted that the x-axis, y-axis and z-axis in this invention are the three coordinate axes of the three-dimensional rectangular coordinate system, that is, the three-dimensional ground-penetrating radar device moves along the x-axis direction of the three-dimensional rectangular coordinate system.

[0074] By processing the three-dimensional radar signal data, the three-dimensional radar signal data is converted into three-dimensional radar data grayscale blocks;

[0075] Step 2: Slice the grayscale block of the 3D radar data layer by layer along the z-axis to obtain a sequence of 2D grayscale images of the road to be detected. Input the grayscale images in the sequence of 2D grayscale images of the road to be detected one by one into the trained YOLO 2D disease detection model. The YOLO 2D disease detection model outputs the disease detection result for each 2D grayscale image. The disease detection result is the bounding box of the location of the disease in the 2D grayscale image.

[0076] Calculate the center coordinates of the bounding box at the location of each disease, and record the center coordinates of each bounding box. and center coordinates The z-axis coordinates of the corresponding two-dimensional grayscale image;

[0077] Step 3: Calculate the adaptive neighborhood radius ε0 of the DBSCAN algorithm based on the center coordinates of the bounding boxes of the lesions in each 2D grayscale image, and cluster the lesions in each 2D grayscale image based on the adaptive neighborhood radius ε0.

[0078] Diseases that are clustered together and spatially connected (determined by the z-axis coordinates of the two-dimensional grayscale image) are merged into a three-dimensional disease candidate region (i.e., continuous and spatially related three-dimensional disease candidate regions are extracted based on the clustering results), and the bounding box, center position and size information of each three-dimensional disease candidate region are calculated.

[0079] Step 4: Based on the bounding box, center position, and size information of the three-dimensional defect candidate region, cut out each three-dimensional defect candidate data block from the three-dimensional radar data block of the road to be detected;

[0080] Step 5: Input each 3D disease candidate data block into the trained nnUNet 3D disease segmentation model, perform fine-grained semantic segmentation on each disease candidate region, and obtain the disease boundary from the semantic segmentation.

[0081] Then, connected component analysis and Gaussian filtering are used to process the disease boundaries segmented by semantic segmentation to obtain the processed disease boundaries, thus obtaining complete three-dimensional ground-penetrating radar road hidden disease segmentation data.

[0082] like Figure 1 As shown, the training methods for the YOLO two-dimensional disease detection model and the nnUNet three-dimensional disease segmentation model in this invention are as follows:

[0083] The experimental road was inspected using a 3D ground-penetrating radar (GPR) device to obtain 3D radar data blocks (x0, y0, z0) with a 3D shape of (440, 66, 170). These data blocks were then used for model pre-training and annotation, specifically including the following steps:

[0084] 1) Slice the 3D radar data block layer by layer along the z-axis to generate 2D radar slices. Use the image annotation tool Labellmg to annotate the location and category labels of defects in the 2D slices. Annotated images are shown below. Figure 2 As shown, the annotation results are converted into the input format of the YOLO model, and then the YOLO model is pre-trained using the converted two-dimensional image to obtain the pre-trained weights for two-dimensional disease detection.

[0085] 2) The data block is annotated at the voxel level using the 3D Slicer tool to generate a 3D segmentation mask, such as... Figure 3 As shown, the data is divided into a training set and a validation set, and the model is trained and validated using the training set and the validation set to obtain the pre-trained weights for 3D disease segmentation.

[0086] The trained model is applied in the following ways:

[0087] The three-dimensional data is sliced ​​layer by layer along the z-axis. The pre-trained weight file of the YOLO model is read. The two-dimensional grayscale image sequence is input into the YOLO two-dimensional disease detection model one by one for disease detection. The disease bounding box and slice index are recorded.

[0088] Using the calculated neighborhood radius ε0 = 10 mm and the minimum sample size N min =2 to remove noise points, and then use the spatial center point coordinates and slice index of the two-dimensional detection box as clustering features. The DBSCAN three-dimensional clustering algorithm is used to perform spatial clustering analysis on the detection results to calculate the center position and size information of each disease candidate area.

[0089] Candidate disease areas are cropped from the original 3D radar data blocks. During the cropping process, necessary axis transformation and flipping are performed on the data blocks of the cropped areas, and each data block is converted into a standard NIfTI format file and a standard dataset structure.

[0090] The generated disease area data is input into the nnUNet model for semantic segmentation to obtain accurate disease boundaries, forming a 3D segmentation file of hidden road diseases. Finally, the spatial positioning information of each candidate region is used to merge the segmentation results into complete 3D reconstruction data. The slicing effect of the 3D reconstruction data is shown in the figure. Figure 4 As shown, the 3D reconstruction effect is as follows: Figure 5 As shown.

[0091] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the processing of the three-dimensional radar signal data specifically involves:

[0092] The three-dimensional radar signal data is subjected to dynamic compression, data normalization, and grayscale mapping in sequence.

[0093] The other steps and parameters are the same as in Specific Implementation Method 1.

[0094] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that it processes the three-dimensional radar signal data, converting the three-dimensional radar signal data into three-dimensional radar data grayscale blocks; the specific process is as follows:

[0095] Step 11: Dynamically compress the acquired 3D radar signal data using logarithmic transformation.

[0096] P i,j,k =log(1+αD) i,j,k )

[0097] Among them, P i,j,k This represents the radar signal data corresponding to coordinate point (i,j,k) after dynamic compression;

[0098] α represents the signal amplification factor, with a default value of 10. 4 ;

[0099] Step 1 & 2: Calculate the mean value μ of the radar signal data corresponding to each coordinate point after dynamic compression.

[0100]

[0101] Calculate the standard deviation σ of the radar signal data corresponding to each coordinate point after dynamic compression:

[0102]

[0103] The radar signal data corresponding to each coordinate point are normalized according to the mean μ and standard deviation σ:

[0104]

[0105] Among them, M i,j,k This represents the normalized data corresponding to the coordinate point (i,j,k);

[0106] Step 13: Perform grayscale value mapping on the normalized data corresponding to each coordinate point, and use the grayscale value mapping result of coordinate point (i,j,k) as the grayscale value of voxel (i,h,k):

[0107]

[0108] Wherein, min(M) represents the minimum value in the normalized data corresponding to each coordinate point;

[0109] max(M) represents the maximum value in the normalized data corresponding to each coordinate point;

[0110] H i,j,k This represents the grayscale value of voxel (i,j,k).

[0111] Other steps and parameters are the same as in specific implementation method one or two.

[0112] Through the processing of this embodiment, the three-dimensional radar signal data is mapped into a three-dimensional radar data grayscale block within the range of [0, 255].

[0113] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the center coordinates of the bounding box at the location of the disease are:

[0114]

[0115] in, The minimum x-coordinate of the pixel within the bounding box of the location of the a-th disease;

[0116] The maximum x-coordinate of the pixel within the bounding box representing the location of the a-th disease;

[0117] The minimum ordinate of the pixel within the bounding box representing the location of the a-th disease;

[0118] The maximum ordinate of the pixels within the bounding box representing the location of the a-th disease;

[0119] This represents the center coordinates of the bounding box where the a-th disease is located.

[0120] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0121] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that the method for calculating the adaptive neighborhood radius ε0 is as follows:

[0122] The bounding box center coordinates of the location of the a-th lesion in the two-dimensional grayscale image sequence Calculate the center coordinates of the bounding boxes of other lesions in the two-dimensional grayscale image sequence. The distance is calculated by taking the distance from the given distance and then sorting the calculated distances in ascending order, resulting in the distance vector d. (k) for:

[0123]

[0124] in, Represents the distance sequence d (k) The 1st, 2nd, ..., N'th distance in the sequence;

[0125] N' represents the total number of diseases other than the a-th disease in the two-dimensional grayscale image sequence;

[0126] Based on the distance vector d (k) Get N' pairs of points (x n ,y n ), n=1,2,…,N':

[0127] x n =n,

[0128] Then based on N' pairs of points (x n ,y n Obtain the cumulative sum function curve, and then obtain the adaptive neighborhood radius based on the cumulative sum function curve:

[0129]

[0130] Where (x'(n),y'(n)) represents a pair of points (x... n ,y n The first derivative of the cumulative sum function curve, (x)(n), y)(n) represents the pair of points (x...). n ,y n The value of the second derivative on the cumulative sum function curve.

[0131] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0132] The adaptive neighborhood radius ε0 calculated according to the present invention and the set minimum sample number N minThe DBSCAN clustering algorithm can be executed to perform spatial clustering analysis on the clustering features of two-dimensional grayscale image sequences.

[0133] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that the specific process of connected component analysis is as follows:

[0134] Calculate the number of voxels in each connected disease region, and denote the number of voxels in the i'-th connected disease region as V. i’ Then, diseased areas with voxel counts less than a threshold are removed to obtain the remaining diseased areas.

[0135] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0136] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that the specific process of the Gaussian filtering method is as follows:

[0137] Gaussian filtering was used to smooth the boundary voxels of the remaining diseased areas.

[0138]

[0139] Where (x,y,z) represents the boundary voxels of the remaining diseased area;

[0140] H disease (x+p,y+q,z+r) represents the gray value of voxel (x+p,y+q,z+r) before smoothing.

[0141] This represents the Gaussian function value at position (p, q, r).

[0142] H smooth (x,y,z) represents the gray value of the boundary voxel (x,y,z) in the remaining diseased area after smoothing.

[0143] p, q, r represent the values ​​of the Gaussian function. The offset in the data is the area of ​​overlap between each voxel of the filter and the input data, where p,q,r∈[-3σ0,3σ0], and σ0 represents the standard deviation of the Gaussian distribution.

[0144] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0145] Processing can reduce edge discontinuities in 3D reconstructed data blocks.

[0146] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One to Seven in that the Gaussian function value at position (p,q,r) is:

[0147]

[0148] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0149] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for identifying, locating, and segmenting hidden road defects based on three-dimensional ground-penetrating radar, characterized in that, The method specifically includes the following steps: Step 1: Use a 3D ground-penetrating radar (GPR) device to perform 3D underground space information detection on the road to be inspected, obtain 3D radar signal data, and set the coordinate points... The corresponding radar signal data is denoted as ; in, , , The 3D ground-penetrating radar device moves in a straight line in the horizontal direction. The direction of movement of the 3D ground-penetrating radar device is taken as... Axial direction, This indicates the maximum moving distance of the radar. Axial coordinates, Indicates that the radar is perpendicular to The maximum detection width corresponding to the axial direction Axial coordinates, This represents the maximum depth at which radar electromagnetic waves can propagate underground. Axis direction coordinates; By processing the three-dimensional radar signal data, the three-dimensional radar signal data is converted into three-dimensional radar data grayscale blocks; The specific processing of the three-dimensional radar signal data involves: performing dynamic compression, data normalization, and grayscale value mapping on the three-dimensional radar signal data in sequence. The process involves processing the three-dimensional radar signal data to convert it into three-dimensional radar data grayscale blocks; the specific process is as follows: Step 11: Dynamically compress the acquired 3D radar signal data using logarithmic transformation. in, This indicates the coordinates after dynamic compression. The corresponding radar signal data; Indicates the signal amplification factor; Steps 1 and 2: Calculate the mean value of the radar signal data corresponding to each coordinate point after dynamic compression. : Calculate the standard deviation of the radar signal data corresponding to each coordinate point after dynamic compression. : According to the mean and standard deviation The radar signal data corresponding to each coordinate point are normalized separately: in, Represents coordinate points The corresponding normalized data; Step 13: Map the normalized data for each coordinate point to grayscale values, thus mapping the coordinate points... The grayscale value mapping result is used as a voxel Grayscale value: in, This represents the minimum value in the normalized data corresponding to each coordinate point; This represents the maximum value in the normalized data corresponding to each coordinate point; Voxel representation grayscale value; Step 2: Along the grayscale blocks of the 3D radar data The road to be detected is sliced ​​layer by layer along the axis to obtain a sequence of two-dimensional grayscale images. The grayscale images in the sequence of two-dimensional grayscale images of the road to be detected are input one by one into the trained YOLO two-dimensional disease detection model. The YOLO two-dimensional disease detection model outputs the disease detection result for each two-dimensional grayscale image. The disease detection result is the bounding box of the location of the disease in the two-dimensional grayscale image. Calculate the center coordinates of the bounding box at the location of each disease, and record the center coordinates of each bounding box. and center coordinates The z-axis coordinates of the corresponding two-dimensional grayscale image; Step 3: Calculate the adaptive neighborhood radius of the DBSCAN algorithm based on the center coordinates of the bounding boxes of the lesion locations in each 2D grayscale image. Based on adaptive neighborhood radius Cluster the diseases in each two-dimensional grayscale image; The adaptive neighborhood radius The calculation method is as follows: For the 2D grayscale image sequence, the th The coordinates of the center of the bounding box where the disease is located ( Calculate the center coordinates of the bounding boxes of other lesions in the two-dimensional grayscale image sequence and ( The distance is calculated by taking the distance and then sorting the calculated distances in ascending order, resulting in a distance vector. for: in, Distance sequence The first one in One distance; This indicates that in a two-dimensional grayscale image sequence, except for the first... Total number of diseases other than the primary disease; Based on the distance vector get Point pair , =1 : Then according to Point pair Obtain the cumulative sum function curve, and then derive the adaptive neighborhood radius based on the cumulative sum function curve: in, Represents point pairs The value of the first derivative on the cumulative sum function curve. Represents point pairs The value of the second derivative on the cumulative sum function curve; Patches that are clustered together and spatially connected are merged into a three-dimensional disease candidate region, and the bounding box, center position and size information of each three-dimensional disease candidate region are calculated. Step 4: Based on the bounding box, center position, and size information of the three-dimensional defect candidate region, cut out each three-dimensional defect candidate data block from the three-dimensional radar data block of the road to be detected; Step 5: Input each 3D disease candidate data block into the trained nnUNet 3D disease segmentation model, perform semantic segmentation on each disease candidate region, and obtain the disease boundary obtained from the semantic segmentation. Then, connected component analysis and Gaussian filtering are used to process the disease boundaries extracted by semantic segmentation to obtain the processed disease boundaries.

2. The method for identifying, locating, and segmenting hidden road defects based on three-dimensional ground-penetrating radar according to claim 1, characterized in that, The center coordinates of the bounding box where the disease is located are: in, Indicates the first The minimum x-coordinate of the pixel within the bounding box of the location of the disease; Indicates the first The maximum x-coordinate of the pixels within the bounding box of the location of each disease; Indicates the first The minimum ordinate of the pixel within the bounding box of the location of the disease; Indicates the first The maximum ordinate of the pixels within the bounding box of the location of each disease; ( Indicates the first The center coordinates of the bounding box where each disease is located.

3. The method for identifying, locating, and segmenting hidden road defects based on three-dimensional ground-penetrating radar according to claim 2, characterized in that, The specific process of connected component analysis is as follows: Calculate the number of voxels in each connected disease region, and then... The number of voxels in a connected disease region is denoted as Then, diseased areas with voxel counts less than a threshold are removed to obtain the remaining diseased areas.

4. The method for identifying, locating, and segmenting hidden road defects based on three-dimensional ground-penetrating radar according to claim 3, characterized in that, The specific process of the Gaussian filtering method is as follows: Gaussian filtering was used to smooth the boundary voxels of the remaining diseased areas. in, Voxels representing the boundary of the remaining diseased area; Indicates voxels before smoothing. grayscale value; express The Gaussian function value of the location; This indicates the boundary voxels within the remaining diseased area after smoothing. grayscale value; Represents the Gaussian function value The offset in , This represents the standard deviation of the Gaussian distribution.

5. The method for identifying, locating, and segmenting hidden road defects based on three-dimensional ground-penetrating radar according to claim 4, characterized in that, The The Gaussian function value of the position is: 。

Images