Method and system for automatic segmentation and measurement of road surface diseases based on three-dimensional laser point cloud and deep neural network

Abstract

The present application relates to a kind of based on three-dimensional laser point cloud and the automatic segmentation and measurement method and system of pavement disease of deep neural network, including point cloud data acquisition, point cloud data is converted to image data, making disease segmentation dataset, detection model construction and training and based on the spatial form measurement of pavement disease of detection result;While converting, labeling, segmentation, statistical method is encapsulated and merged, obtains a input point cloud data, and the automatic segmentation measurement system of pavement disease of output disease detection result image and spatial form parameter.This application is based on laser point cloud data, realizes from laser point cloud original data to the whole process automation and intelligentization of identification result, is conducive to improving the efficiency of asphalt pavement disease detection.

Description

Technical Field

[0001] This invention relates to an automatic segmentation and measurement method and system for road surface defects based on three-dimensional laser point clouds and deep neural networks, belonging to the field of civil engineering inspection technology. Background Technology

[0002] Detection of road surface defects is a crucial step in pavement performance assessment and a vital basis for subsequent road maintenance decisions. However, using on-site investigation methods requires closing traffic and is subject to drawbacks such as significant subjective influence and low efficiency.

[0003] To circumvent the aforementioned problems, existing technologies employ road image analysis, but this method still suffers from several issues, such as low image data quality, imperfect algorithms during analysis, and room for improvement in accuracy. Therefore, 3D laser point cloud analysis is integrated into the detection process. As an advanced measurement method, it offers advantages such as speed, real-time processing, high density, high precision, dynamic capabilities, and active operation. Specifically, mounting a 3D laser scanner and its supporting equipment on a vehicle creates a mobile measurement system. This system can rapidly acquire high-resolution 3D laser point cloud data of the road without disrupting traffic. Furthermore, this measurement system is characterized by low cost, low manpower requirements, high safety, and minimal interference from the working environment.

[0004] With the development of technology, users have found that the detection method based on fused 3D laser point clouds, while containing a large amount of rich information, also increases the difficulty of interpreting the point cloud data. Therefore, there is an urgent need to provide a measurement method based on 3D laser point clouds to improve measurement accuracy and efficiency. Summary of the Invention

[0005] This invention provides an automatic segmentation and measurement method and system for pavement defects based on three-dimensional laser point clouds and deep neural networks. Based on laser point cloud data, the automation and intelligence of the entire process from raw laser point cloud data to recognition results is beneficial to improving the efficiency of asphalt pavement defect detection.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] An automatic pavement distress segmentation and measurement method based on 3D laser point cloud and deep neural network includes the following steps:

[0008] Step S1: Start the 3D laser scanner to scan the road surface lanes. The travel encoder triggers the 3D laser scanner to save the 3D scan results of W meters of road surface lanes. The saved 3D scan results are point cloud data files in .dat format.

[0009] Step S2: Convert the point cloud data file into a W×H matrix, where W is the vehicle travel length covered by the point cloud data file and H is the scan width covered by the point cloud data file; at the same time, the height value Z measured by the laser point cloud within the W×H matrix is ​​stored in the corresponding position within the matrix.

[0010] Step S3: Use the maximum normalization method to normalize the height value Z to the range of 0-255 to obtain a grayscale image of the relative height value of the point with size W×H;

[0011] Step S4: Use Labelme software to perform pixel-level annotation on the grayscale image of the relative height values ​​of the points, and convert the annotation results into label images; divide the label images into n parts, determine the proportion of the target pixels in each part, and preset the proportion value;

[0012] If the proportion of the target pixel in the small portion exceeds the preset proportion value, then retain the small portion that meets the requirements and the corresponding grayscale image of the relative height value of the point, and continue to the next step; if the proportion of the target pixel in the small portion does not exceed the preset proportion value, then continue to compare the proportion of the target pixel in the next small portion with the preset proportion value.

[0013] Step S5: Collect the retained small portions that meet the requirements and the corresponding grayscale images of the relative height values ​​of the points to form a dataset, and randomly divide the dataset into a training dataset and a validation dataset.

[0014] Step S6: Load the training dataset and compile the detection model. Train the training dataset by constructing a convolutional neural network, and set the loss function to classification cross-entropy. The output training results are automatically saved in the model_path directory.

[0015] Step S7: Load the validation dataset and the detection model trained in step S6 to obtain the measurement standard IoU of image segmentation. If the measurement standard IoU is greater than 70%, the detection model can be applied. If the measurement standard IoU is less than 70%, repeat steps S1-S6 until the measurement standard IoU of image segmentation is greater than 70%.

[0016] Step S8: Based on the detection model verified in Step S7, sum the segmentation results of the label image at the pixel size to obtain the length l and width d of the road surface defect; according to the generation logic of the grayscale image of the relative height value of the point, calculate the actual average height of the edge pixels of the road surface defect in reverse. Height of the lowest pixel within the segmentation region of the label image The depth h of the road surface distress was obtained;

[0017] Step S9: Convert the point cloud data files from steps S2-S3 into image data, create a pavement distress segmentation dataset in steps S4-S5, construct and train the detection model in step S6, verify the detection model in step S7, and measure, encapsulate, and merge the actual spatial morphology of pavement distress based on the detection model in step S8. Input the point cloud data files collected in step S1 and output the distress detection result image and spatial morphology parameters of the pavement distress automatic segmentation and measurement system.

[0018] As a further preferred embodiment of the present invention, in step S1, the device for acquiring point cloud data files includes an RTK, a 3D laser scanner, and a traveling encoder, wherein the RTK is installed on the top of the vehicle and is used to acquire the latitude and longitude position of the vehicle in real time.

[0019] A fixed bracket is installed at the rear of the vehicle, and a 3D laser scanner is mounted on the fixed bracket. The scanning port of the 3D laser scanner is perpendicular to the road surface.

[0020] The travel encoder is installed at the left rear wheel position of the vehicle;

[0021] As a further preferred embodiment of the present invention, in step S4, the preset ratio value is defined as η, and its range is 10-25;

[0022] As a further preferred embodiment of the present invention, in step S5, the dataset is randomly divided into a training dataset and a validation dataset, and the ratio of the training dataset to the validation dataset is n1:n2, and n1:n2 = 9:1 or n1:n2 = 8:2;

[0023] As a further preferred embodiment of the present invention, in step S6, when compiling the detection model, the training dataset is loaded, data augmentation and one-hot label format conversion are performed on the training dataset, a convolutional neural network is constructed based on tensorflow-gpu, and the training parameters are set in the convolutional neural network as follows: BATCH_SIZE=b, EPOCHS=p, LEARNING_RATE=α, BUFFER_SIZE=t, model_path=str1, and the loss function is output.

[0024] As a further preferred embodiment of the present invention, in step S6, the output loss function is set to the classification cross-entropy, i.e.

[0025]

[0026] As a further preferred embodiment of the present invention, in step S7, the measurement standard IoU for image segmentation is obtained as follows:

[0027]

[0028] The automatic road surface segmentation and measurement system generated according to the automatic road surface segmentation and measurement method based on three-dimensional laser point cloud and deep neural network is a program that integrates point cloud data files into image data, creates road surface segmentation datasets, constructs and trains detection models, verifies detection models, and measures the actual spatial morphology of road surface defects based on the detection models.

[0029] The 3D laser scanner saves the 3D scanning results of the road surface, i.e., the point cloud data file, which is then input into the integrated program to output the image of the road defect detection result.

[0030] By employing the above technical solutions, the present invention has the following beneficial effects compared to the prior art:

[0031] 1. The automatic segmentation and measurement method for road surface defects based on three-dimensional laser point cloud and deep neural network provided by the present invention uses a three-dimensional laser scanner to scan the road surface and driving lane, which can quickly and with high resolution obtain the three-dimensional scanning results of the road surface. At the same time, the three-dimensional laser scanner has an automatic distortion correction function, providing accurate data of defects for the spatial morphology measurement of road surface defects based on the detection results.

[0032] 2. The automatic segmentation and measurement method for road surface defects based on three-dimensional laser point cloud and deep neural network provided by this invention uses classification cross-entropy as the result of the loss function when constructing and training the detection model, which shortens the acquisition time and ensures high accuracy of the result.

[0033] 3. The automatic segmentation and measurement method for road surface defects based on three-dimensional laser point cloud and deep neural network provided by the present invention can obtain the specific length, width and depth of the defects by measuring the actual spatial morphology of the road surface defects based on the detection results.

[0034] 4. The automatic segmentation and measurement method for pavement defects based on three-dimensional laser point cloud and deep neural network provided by this invention realizes the automation and intelligence of the entire process from the original laser point cloud data to the recognition result, which is conducive to improving the efficiency of asphalt pavement defect detection. Attached Figure Description

[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0036] Figure 1 This is a grayscale image of the relative height values ​​of the points obtained in a preferred embodiment of the present invention;

[0037] Figure 2 These are label images obtained in a preferred embodiment of the present invention;

[0038] Figure 3This is a schematic diagram of the result of segmenting the label image to increase the proportion of the target data in a preferred embodiment provided by the present invention;

[0039] Figure 4 This is a schematic diagram of the training results output in a preferred embodiment of the present invention;

[0040] Figure 5 This is a verification illustration of multiple sets of grayscale images containing relative height values ​​of points, label images, and segmentation results of detection models in a preferred embodiment of the present invention.

[0041] Figure 6 This is the measurement result output from the spatial morphology measurement of pavement defects based on the detection results in the preferred embodiment of the present invention. Detailed Implementation

[0042] The present invention will now be described in further detail with reference to the accompanying drawings. The specific dimensions used in this embodiment are merely illustrative and do not limit the scope of protection of the present invention.

[0043] As described in the background section, under the premise of advanced measurement methods based on three-dimensional laser point clouds, the large amount of point cloud data contained therein leads to increased interpretation, which affects the output efficiency and reduces the accuracy of the results. Therefore, this application provides an automatic segmentation and measurement method for pavement defects based on three-dimensional laser point clouds and deep neural networks, realizing pavement defect identification based on laser point cloud data. Moreover, this technology automates and intelligently transforms the entire process from raw laser point cloud data to identification results, thereby improving the efficiency of asphalt pavement defect detection.

[0044] Before detailing the detection method, the measurement system used in this application, including an RTK scanner, a 3D laser scanner, and a traveling encoder, is provided, all integrated on the vehicle. The RTK is mounted on the top of the vehicle to acquire its latitude and longitude position in real time. A fixed bracket is installed at the rear of the vehicle, on which the 3D laser scanner is mounted. The scanning port of the 3D laser scanner is perpendicular to the road surface, and the distance between the 3D laser scanner and the road surface is determined based on the scanning range of the 3D laser scanner, ensuring that the scanning range covers at least one lane of the road. The traveling encoder is installed at the left rear wheel of the vehicle to accurately measure the vehicle's travel distance and control the data storage of the 3D laser scanner.

[0045] The following section details the specific steps of the automatic pavement distress segmentation and measurement method based on 3D laser point clouds and deep neural networks provided in this application:

[0046] Step S1: Start the 3D laser scanner to scan the road surface lane. Every W meters the vehicle travels, the travel encoder triggers the 3D laser scanner to save the 3D scan result of the W meters of road surface lane. The aforementioned saved 3D scan result of the road surface is a point cloud data file in .dat format.

[0047] Step S2: Using a point cloud data file in .dat format as input, convert the point cloud data file into a W×H matrix, where W is the vehicle travel length covered by the point cloud data file and H is the scan width covered by the point cloud data file; at the same time, the height value Z measured by the laser point cloud within the W×H matrix is ​​stored in the corresponding position within the matrix; the program input here is actually an automatic conversion program that can decode the point cloud data file into a W×H matrix.

[0048] Step S3: Use the maximum normalization method to normalize the height value Z to the range of 0-255 to obtain a grayscale image of the relative height value of the point with size W×H.

[0049] Step S4: Use Labelme software to perform pixel-level annotation on the grayscale image of the relative height values ​​of the points, and convert the annotation results into label images; since most of the pixels in the label image are background pixels (i.e., unlabeled areas), this will affect the training results. Therefore, the label image is divided into n parts, and the proportion of the target pixels in each part is determined. The preset proportion value is defined as η, which is generally in the range of 10-25.

[0050] If the proportion of the target pixel in the small portion exceeds the preset proportion value, then retain the small portion that meets the requirements and the corresponding grayscale image of the relative height value of the point, and continue to the next step; if the proportion of the target pixel in the small portion does not exceed the preset proportion value, then continue to compare the proportion of the target pixel in the next small portion with the preset proportion value.

[0051] Step S5: Collect the retained small portions that meet the requirements and the corresponding grayscale images of relative height values ​​of the points to form a dataset, and randomly divide the dataset into a training dataset and a validation dataset; the ratio of the training dataset to the validation dataset is n1:n2, and n1:n2 = 9:1 or n1:n2 = 8:2.

[0052] Step S6: Load the training dataset, compile the detection model, perform data augmentation and "one-hot" label format conversion on the training dataset, build a convolutional neural network based on TensorFlow-GPU, and set the training parameters in the convolutional neural network as follows: BATCH_SIZE=b, EPOCHS=p, LEARNING_RATE=α, BUFFER_SIZE=t, model_path=str1. The obtained loss function is set as classification cross-entropy. There are actually various forms of loss function representation, but after verification, it was found that when classification cross-entropy is used as the result, the acquisition time is the shortest and the accuracy is the highest. Among these:

[0053] Categorical cross-entropy is...

[0054]

[0055] Compile the detection model and start training. The output training results will be automatically saved in the model_path directory.

[0056] Step S7: Load the validation dataset and the detection model trained in Step S6 to obtain the IoU measurement standard for image segmentation. IoU is...

[0057]

[0058] If the measured standard IoU exceeds 70%, the detection model can be applied. If the measured standard IoU is below 70%, the dataset needs to be enriched, and steps S1-S6 are repeated until the measured standard IoU of image segmentation exceeds 70%.

[0059] Step S8: Based on the detection model verified in Step S7, sum the segmentation results of the label image at the pixel size to obtain the length l and width d of the road surface defect; according to the generation logic of the grayscale image of the relative height value of the point, calculate the actual average height of the edge pixels of the road surface defect in reverse. Height of the lowest pixel within the segmentation region of the label image The depth h of the road surface distress was obtained;

[0060] It is important to note that this application uses a 3D laser scanner not only because it can quickly and at high resolution acquire 3D laser point cloud data of roads, but also because the 3D laser scanner has an automatic distortion correction function, which is not available in ordinary scanners. In the W×H point matrix obtained by scanning and decoding the image by the 3D laser scanner, the horizontal and vertical spacing between each point and the surrounding points is equal to t, that is, the actual distance between two adjacent pixels in the obtained image is also t. From this, the actual length, width and depth of the road surface defects can be obtained, and the obtained actual length, width and depth are also relatively accurate.

[0061] Step S9: Convert the point cloud data files from steps S2-S3 into image data; create a pavement distress segmentation dataset in steps S4-S5; construct and train the detection model in step S6; verify the detection model in step S7; and perform spatial morphological measurement, encapsulation, and merging of pavement distress based on the detection model in step S8. Input the point cloud data files collected in step S1 and output the distress detection result image and spatial morphological parameters of the pavement distress automatic segmentation and measurement system.

[0062] The final automatic pavement distress segmentation and measurement system generated by the above-mentioned method based on 3D laser point cloud and deep neural network is a program that integrates point cloud data files into image data, creates pavement distress segmentation datasets, constructs and trains detection models, verifies detection models, and measures the actual spatial morphology of pavement distress based on the detection models. The 3D scanning results of the pavement surface saved by the 3D laser scanner, i.e., the point cloud data file, are input into the integrated program, and the distress detection result image is output.

[0063] Example

[0064] To verify the superiority of the above-mentioned automatic pavement distress segmentation and measurement method and system, the implementers chose a clear day to calibrate the 3D laser scanner; they drove the inspection vehicle for 3km on Jiyin Avenue in Nanjing and obtained 680 valid .dat point cloud data files.

[0065] Taking 680 valid DAT point cloud data files as input, the program decodes them into a matrix of size 3639×2665×Z'. Each point cloud data file covers a vehicle travel length of 3639 pixels and a scan width of 2665 pixels. i The height is a value measured from a laser point cloud within a range of 3639×2665; the automatic conversion program uses the maximum normalization method to convert these height values ​​Z... i Normalizing to the range of 0-255, the result is as shown in the attached figure. Figure 1 The image shows a grayscale representation of the relative height values ​​of points with dimensions of 3639×2665.

[0066] The Labelme software was used to perform pixel-level annotation on the grayscale image of the relative height values ​​of the points, and the annotation results are shown in the attached figure. Figure 2 The labeled image shown is typical; most pixels in a labeled image are background pixels (i.e., unlabeled areas), which can affect training results. Therefore, a labeled image is divided into 24 parts, and the proportion of target pixels in each part is determined. If this proportion exceeds 1%, that part and its corresponding original image are retained. A schematic diagram is attached. Figure 3 As shown, the dataset is formed in this way; the dataset is randomly divided into a training dataset and a validation dataset, with a ratio of 9:1.

[0067] In the convolutional neural network, set the training parameters as follows: BATCH_SIZE=9, EPOCHS=300, LEARNING_RATE=0.0001, BUFFER_SIZE=30, model_path='. / w-net.h5', and set the loss function to Categorical Crosssentropy. Then compile the detection model and start training. The training results will be automatically saved in the working directory and named w-net.h5. The training output is shown in the attached image. Figure 4 ; Figure 4 In the diagram, ① represents Loss and ② represents IoU. During the 300 iterations, it is evident that the loss value first decreases rapidly and then stabilizes at a low level, while the IoU ratio first increases steadily and then stabilizes at a high level.

[0068] Repeat steps S1-S6, adjusting parameters as needed, until the verified IoU reaches over 70%. Here is an example... Figure 5 The diagrams (5a-5f) show multiple sets of comparison results. In each set of diagrams, the leftmost column "realimage" is the original grayscale image data of the input model, the middle column "reallabel" is the result of manual annotation, and the rightmost column "predlabel" is the model segmentation result. It can be seen intuitively that the segmentation result obtained based on the method provided in this application is good, and the average IoU of the validation set is 0.7204.

[0069] After verifying the feasibility of the detection model, the 3D laser scanner used in this application can obtain a 3639×2665 dot matrix where the horizontal and vertical spacing between each point and its surrounding points is equal to 1.2mm, meaning the actual distance between two adjacent pixels in the resulting image is also 1.2mm. Summing the segmentation results over the pixel dimensions yields a lesion length of 1800mm and a width of 2.4mm. Based on the generation logic of the grayscale image of the relative height values ​​of the points, the lesion depth is calculated to be 7.2mm. The final accurate output results are attached. Figure 6 As shown.

[0070] This application encapsulates and merges the aforementioned transformation, annotation, segmentation, and statistical methods to obtain an automatic pavement distress segmentation and measurement system that outputs not only distress detection result images but also spatial morphological parameters from a single input point cloud data set. This system automates and intelligently integrates the entire process of laser point cloud technology from data acquisition to pavement distress measurement, reducing labor costs and processing time.

[0071] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0072] The meaning of "and / or" as used in this application includes situations where each exists alone or both exist simultaneously.

[0073] The term "connection" as used in this application can mean a direct connection between components or an indirect connection between components through other components.

[0074] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. An automatic segmentation and measurement method for road surface defects based on three-dimensional laser point clouds and deep neural networks, characterized in that: Specifically, the following steps are included: Step S1: Start the 3D laser scanner to scan the road surface lanes. The travel encoder triggers the 3D laser scanner to save the 3D scan results of W meters of road surface lanes. The saved 3D scan results are point cloud data files in .dat format. Step S2: Convert the point cloud data file into a W×H matrix, where W is the vehicle travel length covered by the point cloud data file and H is the scan width covered by the point cloud data file; at the same time, the height value Z measured by the laser point cloud within the W×H matrix is ​​stored in the corresponding position within the matrix. Step S3: Use the maximum normalization method to normalize the height value Z to the range of 0-255 to obtain a grayscale image of the relative height value of the point with size W×H; Step S4: Use Labelme software to perform pixel-level annotation on the grayscale image of the relative height values ​​of the points, and convert the annotation results into label images; divide the label images into n parts, determine the proportion of the target pixels in each part, and preset the proportion value; If the proportion of the target pixel in the small portion exceeds the preset proportion value, then retain the small portion that meets the requirements and the corresponding grayscale image of the relative height value of the point, and continue to the next step; if the proportion of the target pixel in the small portion does not exceed the preset proportion value, then continue to compare the proportion of the target pixel in the next small portion with the preset proportion value. Step S5: Collect the retained small portions that meet the requirements and the corresponding grayscale images of the relative height values ​​of the points to form a dataset, and randomly divide the dataset into a training dataset and a validation dataset. Step S6: Load the training dataset and compile the detection model. Train the training dataset by constructing a convolutional neural network, and set the loss function to classification cross-entropy. The output training results are automatically saved in the model_path directory. Step S7: Load the validation dataset and the detection model trained in step S6 to obtain the measurement standard IoU of image segmentation. If the measurement standard IoU is greater than 70%, the detection model can be applied. If the measurement standard IoU is less than 70%, repeat steps S1-S6 until the measurement standard IoU of image segmentation is greater than 70%. Step S8: Based on the detection model verified in Step S7, sum the segmentation results of the label image at the pixel size to obtain the length l and width d of the road surface defect; according to the generation logic of the grayscale image of the relative height value of the point, calculate the actual average height of the edge pixels of the road surface defect in reverse. Height of the lowest pixel within the segmentation region of the label image The depth h of the road surface distress was obtained; Step S9: Convert the point cloud data files from steps S2-S3 into image data; create a pavement distress segmentation dataset in steps S4-S5; construct and train the detection model in step S6; verify the detection model in step S7; and perform spatial morphological measurement, encapsulation, and merging of pavement distress based on the detection model in step S8. Input the point cloud data files collected in step S1 and output the distress detection result image and spatial morphological parameters of the pavement distress automatic segmentation and measurement system.

2. The method for automatic segmentation and measurement of road surface defects based on three-dimensional laser point cloud and deep neural network according to claim 1, characterized in that: In step S1, the devices for acquiring point cloud data files include RTK, 3D laser scanner and traveling encoder. The RTK is installed on the top of the vehicle and is used to acquire the vehicle's latitude and longitude position in real time. A fixed bracket is installed at the rear of the vehicle, and a 3D laser scanner is mounted on the fixed bracket. The scanning port of the 3D laser scanner is perpendicular to the road surface. The travel encoder is installed at the left rear wheel position of the vehicle.

3. The method for automatic segmentation and measurement of road surface defects based on three-dimensional laser point cloud and deep neural network according to claim 2, characterized in that: In step S4, the preset ratio value is defined as η, which ranges from 10 to 25.

4. The method for automatic segmentation and measurement of road surface defects based on three-dimensional laser point cloud and deep neural network according to claim 3, characterized in that: In step S5, the dataset is randomly divided into a training dataset and a validation dataset. The ratio of the training dataset to the validation dataset is n1:n2, and n1:n2 = 9:1 or n1:n2 = 8:

2.

5. The method for automatic segmentation and measurement of road surface defects based on three-dimensional laser point cloud and deep neural network according to claim 4, characterized in that: In step S6, when compiling the detection model, the training dataset is loaded, and data augmentation and one-hot label format conversion are performed on the training dataset. A convolutional neural network is built based on tensorflow-gpu. The training parameters in the convolutional neural network are set as follows: BATCH_SIZE=b, EPOCHS=p, LEARNING_RATE=α, BUFFER_SIZE=t, model_path=str1, and the loss function is output.

6. The method for automatic segmentation and measurement of road surface defects based on three-dimensional laser point cloud and deep neural network according to claim 5, characterized in that: In step S6, the output loss function is set to the classification cross-entropy, i.e.

7. The method for automatic segmentation and measurement of road surface defects based on three-dimensional laser point cloud and deep neural network according to claim 6, characterized in that: In step S7, the IoU measurement standard for image segmentation is obtained as follows:

8. The automatic pavement distress segmentation and measurement system generated by the automatic pavement distress segmentation and measurement method based on three-dimensional laser point cloud and deep neural network according to claim 7, characterized in that: The automatic pavement distress segmentation and measurement system is a program that integrates point cloud data files into image data, creates pavement distress segmentation datasets, constructs and trains detection models, verifies detection models, and measures the actual spatial morphology of pavement distresses based on the detection models. The 3D laser scanner saves the 3D scanning results of the road surface, i.e., the point cloud data file, which is then input into the integrated program to output the image of the road defect detection result.

Images