Sports runway pavement damage identification method

Abstract

The invention discloses a sports runway pavement damage identification method. The method comprises the following steps: establishing a construction defect template and a use wear template; controlling the unmanned aerial vehicle to fly along the geometrical shape of the runway through a preset flight path, adopting automatic exposure adjustment and shutter speed optimization to adapt to cloudy or strong light conditions, and collecting high-resolution runway surface images with the pixel resolution reaching the micron level to obtain a first image set; for the first image set, smoothing light interference noise by adopting bilateral filtering processing and reserving edge details, and correcting the influence of strong light reflection or shadow through pixel value normalization processing to obtain a second image set; and extracting defect features from the second image set, performing multi-scale feature analysis by adopting a convolutional neural network, calculating a crack length, a recess area and a bump height, generating a high-dimensional feature vector containing defect shapes and textures, and obtaining a defect feature set.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method for identifying pavement damage of a sports track. Background Art

[0002] The application of drones in the field of rubber track inspection in stadiums is of great significance, especially in ensuring the safety and maintenance quality of sports venues. By using drones equipped with high-resolution cameras to regularly photograph the surface of the track and combining image processing technology to analyze the type of damage, it is not only possible to effectively monitor the status of the track, but also to provide a scientific basis for liability determination. The widespread application of this technology helps to extend the service life of the track, reduce maintenance costs, and ensure the safety of athletes. However, existing inspection methods mainly rely on manual inspections or fixed camera monitoring, which have the defects of limited coverage, low efficiency, and difficulty in capturing subtle damage. Manual inspection is limited by the resolution of the naked eye, making it difficult to detect tiny cracks or slight bulges at an early stage, while fixed cameras are limited by the installation position and cannot achieve high-precision scanning of the entire track. These limitations make it difficult to achieve early warning and accurate classification of track damage, which in turn affects the accurate distinction between construction quality problems and wear and tear. Summary of the invention

[0003] The purpose of the present invention is to solve the above-mentioned problems and provide a method for identifying pavement damage of a sports track.

[0004] The technical solution of the present invention is achieved in this way: The present invention provides a method for identifying sports track pavement damage, the method comprising: Establishing construction defect templates and using worn templates; The drone is controlled to fly along the runway geometry through a preset flight path, and automatic exposure adjustment and shutter speed optimization are used to adapt to cloudy or bright light conditions to collect high-resolution runway surface images with a pixel resolution of micrometers to obtain a first image set; For the first image set, bilateral filtering is used to smooth light interference noise and retain edge details, and pixel value normalization is used to correct the influence of strong light reflection or shadow to obtain the second image set; Defect features are extracted from the second image set, and multi-scale feature analysis is performed using a convolutional neural network to calculate the crack length, the concave area, and the bulge height, generate a high-dimensional feature vector containing the defect shape and texture, and obtain a defect feature set; If the similarity between the feature vector in the defect feature set and the construction defect template is higher than the first threshold value T1, it is judged as a construction defect and the first classification result is obtained; otherwise, the subsequent judgment is performed; If the similarity between the feature vector in the defect feature set and the wear template is higher than the second threshold T2, it is judged as wear and tear and a second classification result is obtained; otherwise, it is marked as an unclassified defect and a third classification result is obtained.

[0005] The advantages or beneficial effects of the above technical solution include at least: The present invention discloses an intelligent detection method for runway surface damage based on drones. The method controls drones to collect high-resolution runway surface images through a preset flight path. After image processing and feature extraction, convolutional neural networks and support vector machines are used to classify defects to determine whether they are construction defects or wear and tear. Subsequently, a damage distribution heat map is generated to determine priority maintenance areas and optimize the flight path for the next inspection. The present invention realizes the automated detection, classification and evaluation of runway surface damage, improves detection efficiency and accuracy, provides a scientific basis for runway maintenance, and helps to ensure flight safety and extend the service life of the runway. DETAILED DESCRIPTION

[0006] It should be understood that the term "including" and its variations used in this document are open inclusions, that is, "including but not limited to". The term "based on" means "based at least in part on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments". Relevant definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc. mentioned in the present invention are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0007] It should be noted that the modifications of "one" and "plurality" mentioned in the present invention are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0008] The names of the messages or information exchanged between multiple devices in the embodiments of the present invention are only used for illustrative purposes and are not intended to limit the scope of these messages or information.

[0009] A method for identifying sports track pavement damage, the method comprising: S1. Establish a construction defect template and a wear template, control the UAV to fly along the runway geometry through a preset flight path, use automatic exposure adjustment and shutter speed optimization to adapt to cloudy or strong light conditions, and collect high-resolution runway surface images with a pixel resolution of microns to obtain the first image set.

[0010] The flight trajectory based on the runway geometry is generated by the preset path generation algorithm, and the flight trajectory data is output. The UAV control system is used to perform precise flight according to the flight trajectory data and output the flight status data. Based on the flight status data, the ambient light intensity is obtained through the light sensor to determine the lighting conditions. If the light intensity is lower than the preset threshold, the automatic exposure algorithm is used to adjust the camera parameters and output the cloudy optimization parameters; if the light intensity is higher than the preset threshold, the shutter speed is adjusted and the strong light optimization parameters are output. The camera is configured according to the optimized parameters to collect micron-level resolution runway surface images to obtain the first image set.

[0011] Specifically, the drone is controlled to fly along the runway geometry by a preset flight path, and automatic exposure adjustment and shutter speed optimization are used to adapt to cloudy or strong light conditions, and high-resolution runway surface images with a pixel resolution of micrometers are collected to obtain the first image set. First, the drone flight path planning adopts a path generation algorithm based on B-spline curves to ensure that the drone can accurately fly along the centerline of the runway. The flight altitude is controlled at 10 meters and the flight speed is 2 meters per second to ensure the stability and clarity of image acquisition. During the flight, the camera carried by the drone monitors the ambient light intensity in real time through an automatic exposure adjustment algorithm. When the light intensity is lower than 1000 lux, the exposure time is automatically adjusted to 1 / 100s and the aperture value is set to f / 2.8 to ensure that clear images can still be obtained under cloudy conditions; when the light intensity is higher than 10,000 lux, the exposure time is adjusted to 1 / 1000 seconds and the aperture value is set to f / 8 to avoid overexposure under strong light. At the same time, the shutter speed optimization algorithm sets the shutter speed to 1 / 500 second according to the flight speed and image resolution requirements to reduce motion blur and ensure the clarity of image details. During the image acquisition process, the camera uses a global shutter mode and shoots continuously at a speed of 30 frames per second. The resolution of each frame is 4000×3000 pixels and the pixel size is 1.4 microns, ensuring that the details of the runway surface can be accurately captured. The collected images are denoised and enhanced in real time through the image processing algorithm. The denoising algorithm based on wavelet transform is used to remove Gaussian noise and salt and pepper noise in the image to improve the image quality. Finally, the image stitching algorithm is used to stitch multiple frames of images into a complete runway surface image. The stitching accuracy reaches the sub-pixel level, ensuring the seamless connection of the image and the integrity of the details, forming the first image set.

[0012] S2. For the first image set, bilateral filtering is used to smooth light interference noise and retain edge details, and pixel value normalization is used to correct the influence of strong light reflection or shadow to obtain a second image set.

[0013] Initial pixel data is obtained from the first image set, and bilateral filtering is used to process light interference noise, retain edge details, and obtain an intermediate image set. For the intermediate image set, if the pixel value exceeds the preset strong light threshold, the strong light reflection is corrected by normalizing the pixel value to obtain a corrected image set. For the corrected image set, if a shadow area is detected, local contrast enhancement is used to adjust the shadow area to obtain an enhanced image set. Edge features are extracted from the enhanced image set, and the edge detection algorithm is used to determine the integrity of edge details to obtain an edge feature set. Based on the edge feature set, mean filtering is used to smooth the noise in non-edge areas to obtain a smoothed image set. For the smoothed image set, the overall brightness distribution is optimized by histogram equalization to obtain a second image set. The final pixel data is obtained from the second image set to determine whether it meets the preset image quality standard to obtain a qualified image set.

[0014] In the image processing process, the first image set is smoothed using the bilateral filtering algorithm to eliminate light interference noise. Bilateral filtering combines spatial proximity and pixel value similarity, and effectively retains edge details by setting the spatial domain standard deviation σ_d to 3 and the pixel value domain standard deviation σ_r to 15.

[0015] For example, when processing an image with complex textures, bilateral filtering can smooth the background while maintaining the sharpness of the edges, significantly improving the image quality. Next, pixel value normalization is used to correct the effects of strong light reflections or shadows. The specific operation is to linearly map the image pixel values ​​to the range of 0 to 255, and use the minimum and maximum normalization method to map the pixel values ​​from the original range [50, 200] to [0, 255], thereby eliminating the effects of uneven lighting.

[0016] S3. Extract defect features from the second image set, use convolutional neural network to perform multi-scale feature analysis, calculate crack length, depression area and bulge height, generate a high-dimensional feature vector containing defect shape and texture, and obtain a defect feature set.

[0017] The defect image is obtained from the second image set, and the image preprocessing technology is used to remove the noise to obtain a clear defect image. The multi-scale feature extraction of the clear defect image is performed through the convolutional neural network to generate a multi-scale feature map. For the multi-scale feature map, the crack length, the depression area and the bulge height are calculated to obtain the defect geometric parameters. If the defect geometric parameters exceed the preset threshold, the texture analysis algorithm is used to extract the defect texture features and generate a texture feature set. According to the defect geometric parameters and the texture feature set, a high-dimensional feature vector containing the defect shape and texture is constructed to obtain a feature vector set. The feature vector set is classified by the clustering algorithm to determine the defect feature set. If the defect feature set matches the preset defect template, the feature fusion technology is used to generate the final defect feature set.

[0018] Specifically, when extracting defect features from the second image set, a convolutional neural network is first used for multi-scale feature analysis. By designing a network structure containing multiple convolutional layers, for example, using 3×3 and 5×5 convolution kernels to extract local and global features respectively, the ReLU activation function is combined to enhance the nonlinear expression ability. During the feature extraction process, transfer learning is performed using pre-trained models such as VGG-16 or ResNet-50, and the resolution of the input image is adjusted to 512×512 pixels to ensure that detail information is not lost.

[0019] For the calculation of crack length, the Canny edge detection algorithm is used to extract the crack contour, combined with the Hough transform to fit the straight line segment, and the crack length is calculated through the mapping relationship between pixels and actual sizes (for example, 1 pixel corresponds to 0.1 mm). Assuming that the pixel length of a crack detected is 150, the actual length is 15 mm. For the calculation of the concave area, the image segmentation algorithm (such as U-Net) is used to separate the concave area from the background, and the number of pixels in the concave area is counted. Assuming that the area of ​​the concave area is 3000 pixels, the actual area is 300 square millimeters. For the calculation of the bulge height, the 3D point cloud data of the bulge is obtained through 3D reconstruction technology (such as structured light scanning), and the vertical distance between the bulge vertex and the reference surface is calculated. Assuming that the bulge height is 2.5 mm. Finally, the extracted defect shape is fused with texture information to generate a high-dimensional feature vector. For example, PCA dimensionality reduction is used to compress the 1000-dimensional feature vector to 128 dimensions to form a defect feature set, which provides data support for subsequent defect classification and evaluation.

[0020] S4. If the similarity between the feature vector in the defect feature set and the construction defect template is higher than the first threshold T1 (set to 0.85-0.9, preferably, T1 is set to 0.85), it is judged as a construction defect and the first classification result is obtained. Otherwise, proceed to the subsequent judgment.

[0021] Obtain defect image data, remove noise and adjust contrast through image preprocessing technology to obtain standardized defect images. Extract feature vectors from standardized defect images, use convolutional neural networks to extract deep features, and construct defect feature sets. Calculate the cosine similarity between the defect feature vector and each template in the preset construction defect template library, and record the similarity value corresponding to each template. If the similarity between the defect feature vector and any construction defect template is higher than the first threshold T1, it is determined to be a construction defect and the first classification result is marked. For samples whose first classification results are non-construction defects, calculate the similarity between their feature vectors and the design defect template. If it is higher than the second threshold T2, it is determined to be a design defect. For defect samples whose categories cannot be determined by template matching, secondary classification is performed based on the random forest algorithm, taking into account multi-dimensional features such as defect morphology, location, and material. Generate a defect identification report based on the classification results, including defect type, location information, similarity value and confidence, and associate the corresponding defect handling suggestions.

[0022] Specifically, in the process of construction defect detection, we first need to construct a defect feature set, which contains multiple feature vectors, such as crack width, depth, length, etc. Suppose we have a feature vector of [0.12, 0.08, 0.15], which represents the width, depth and length of the crack. Next, we calculate the similarity between the feature vector and the construction defect template. The feature vector of the preset template is [0.10, 0.07, 0.14], and the cosine similarity algorithm is used for calculation. The formula is cosθ=(A·B) / (||A|| * ||B||), where A is the feature vector and B is the template vector. If the calculated similarity is 0.92, which is higher than the first threshold T1, it is judged as a construction defect and the first classification result is obtained. If the similarity is lower than 0.85, it enters the subsequent judgment and uses other algorithms such as Euclidean distance or Manhattan distance for further analysis to ensure the accuracy of the classification results. In this way, the system can automatically and efficiently identify construction defects and provide a basis for subsequent repair work.

[0023] S5. If the similarity between the feature vector in the defect feature set and the wear template is higher than the second threshold value T2, it is judged as wear and tear and a second classification result is obtained. Otherwise, it is marked as an unclassified defect and a third classification result is obtained.

[0024] Extract defect feature vectors from input data, and generate feature vectors using a preset feature extraction algorithm. Calculate similarity between feature vectors and preset wear templates, and use cosine similarity algorithm to obtain similarity values. If the similarity value exceeds the preset second threshold T2, it is judged as wear due to use, and the second classification result is obtained. If the similarity value does not exceed the preset second threshold T2, it is marked as an unclassified defect, and the third classification result is obtained. For the third classification result, obtain feature vectors of unclassified defects, group them using a clustering algorithm, and obtain defect subcategories. Match defect subcategories with preset extended template libraries to determine potential defect types and obtain extended classification results. According to the extended classification results, update the preset wear template library, and use an incremental learning algorithm to obtain an updated template library.

[0025] In one possible implementation, extracting defect feature vectors is the core of construction defect detection. Feature extraction algorithms are usually based on image processing technology to extract key information from input construction material images.

[0026] For example, the edge detection algorithm can be used to identify the boundaries of cracks, and then generate a feature vector containing crack width, depth, and pixel intensity. Assume that a concrete surface image is processed by the algorithm and a feature vector [0.15, 0.09, 0.20] is extracted, representing the crack width, depth, and pixel intensity, respectively. This method ensures the accuracy of the feature vector and provides a reliable basis for subsequent analysis.

[0027] Specifically, the cosine similarity algorithm is used to calculate the similarity between the feature vector and the preset wear template. The preset wear template stores typical wear features, such as the surface crack template caused by long-term use, whose feature vector is [0.14, 0.08, 0.18]. The cosine similarity calculation is used to compare the closeness between the input feature vector and the template.

[0028] For example, the calculation result shows that the similarity is 0.88, which is higher than the preset second threshold value T2 (assuming it is 0.80), so it is judged as wear and tear, and the second classification result is obtained. This method can quickly distinguish wear and non-wear defects.

[0029] Preferably, if the similarity is lower than T2, for example, if the similarity of a feature vector is 0.75, it is marked as an unclassified defect, and the third classification result is obtained. For these unclassified defects, a clustering algorithm (such as K-means) is used for grouping. Assuming that there are multiple unclassified feature vectors, they are divided into two groups through the clustering algorithm, one group may represent shallow cracks, and the other group represents deep cracks.

[0030] For example, the eigenvectors of the shallow crack group are concentrated around [0.10, 0.05, 0.12]. This grouping helps reveal the underlying pattern of defects.

[0031] In one embodiment, the defect subclassification is matched with a preset extended template library to determine the potential defect type. The extended template library contains a variety of atypical defect templates, such as cracks caused by chemical corrosion, whose feature vector is [0.13, 0.06, 0.15]. Through matching, the shallow crack group may be identified as chemical corrosion defects, and an extended classification result is obtained. This method improves the comprehensiveness of defect identification.

[0032] For example, when updating the preset wear template library, an incremental learning algorithm is used. Based on the extended classification results, the newly identified chemical corrosion defect feature vector is added to the template library. Incremental learning updates the model through small batches of data, avoiding retraining the entire template library.

[0033] The new template library contains updated chemical corrosion templates [0.13, 0.06, 0.15], which improves the adaptability of subsequent inspections. This dynamic update mechanism ensures that the template library matches the actual construction scene.

[0034] The logical progression of the above method, from feature extraction to template update, forms a complete defect detection closed loop. The implementation methods of each link support each other, ensuring the reliability of the classification results and the adaptive ability of the system.

[0035] As a further improvement, in other embodiments, the method further comprises: S6. For the first classification results and the second classification results, a support vector machine is used to map the high-dimensional feature space through a kernel function, optimize the classification boundaries of construction defects and wear and tear, and obtain the final classification results.

[0036] Through kernel function mapping, feature data is obtained from the first classification result and the second classification result, and converted to a high-dimensional feature space to obtain an initial feature set. A support vector machine is used to train a classification model for the initial feature set, optimize the classification boundary of construction defects and wear and tear, and obtain a preliminary classification model. If the classification boundary of the preliminary classification model converges, the classification weight is extracted from the preliminary classification model, and the first classification result and the second classification result are fused to obtain a fused feature set; if not, the kernel function parameters are adjusted and the classification model is retrained. According to the fused feature set, a support vector machine is used for secondary classification, and the final classification boundary of construction defects and wear and tear is determined by the classification boundary optimization algorithm to obtain an optimized classification model. By optimizing the classification model, the classification probability distribution is obtained from the fused feature set, the category attribution of construction defects and wear and tear is determined, and the category prediction result is obtained. According to the category prediction result, a preset threshold is used to judge the classification confidence. If the confidence is higher than the threshold, the final classification result is output; if it is lower than the threshold, abnormal features are extracted from the fused feature set and secondary classification is performed again. Classification features are extracted from the final classification result to generate classification labels for construction defects and wear and tear, and a classification label set is obtained.

[0037] In a possible implementation, kernel function mapping is a process of converting low-dimensional feature data into a high-dimensional feature space to enhance the separability of data.

[0038] For example, in the classification scenario of construction defects and wear and tear, the first classification result may contain the geometric shape features of the defects, such as edge roughness, and the second classification result may contain material wear features, such as surface texture changes. Kernel functions, such as radial basis functions, can map these features to a high-dimensional space, making the originally linearly inseparable features separable.

[0039] For example, assuming that the original feature vector is 2-dimensional, including edge roughness 0.4 and texture change 0.6, after kernel function mapping, a 5-dimensional feature vector may be generated to increase the distinction between features. This mapping provides a richer feature expression for subsequent classification.

[0040] Specifically, the process of training a classification model using a support vector machine aims to find the optimal classification boundary.

[0041] Preferably, the initial feature set includes 1000 samples, including 600 construction defect samples and 400 wear and tear samples. The support vector machine determines the preliminary classification boundary between construction defects and wear and tear by maximizing the interval of the classification boundary.

[0042] For example, training may find that construction defects have higher weights on edge roughness features, while usage wear and tear has higher weights on texture features.

[0043] It should be noted that if the model does not converge, it may be due to improper kernel function parameter settings, such as the radial basis function width parameter is too large, resulting in blurred classification boundaries. In this case, you can retrain by reducing the width parameter until the boundary converges.

[0044] In one embodiment, the generation of the fusion feature set depends on the extraction of classification weights.

[0045] For example, after the initial classification model converges, the extracted weights show that the contribution of edge roughness to construction defects is 0.7, and the contribution of texture change to wear and tear is 0.65. These weights are used to weightedly fuse the first classification result and the second classification result to form a fusion feature set.

[0046] For example, the original features of a sample are edge roughness 0.5 and texture change 0.8. After fusion, a new feature vector may be generated, which comprehensively reflects the characteristics of the two defects. This fusion improves the robustness of classification.

[0047] It can be understood that the secondary classification further improves the classification accuracy by optimizing the classification boundary.

[0048] For example, based on the fused feature set, the support vector machine may generate a finer classification boundary to distinguish construction defects with high edge roughness but low texture variation. The classification probability distribution shows that the probability of a sample belonging to a construction defect is 0.85, and the probability of wear and tear is 0.15.

[0049] Preferably, the confidence threshold is preset to 0.8, and if the confidence of the sample is higher than the threshold, the construction defect label is directly output. If the confidence is lower than 0.8, such as 0.6, abnormal features are extracted, such as an abnormal texture change value of 0.9, and secondary classification is performed again.

[0050] For example, the classification label set of the final classification result is generated based on the extraction of classification features.

[0051] For example, the optimized classification model may identify that the typical characteristics of construction defects are edge roughness greater than 0.5 and texture variation less than 0.3, and generate corresponding classification labels, which can be used for subsequent defect analysis and processing to improve the practicality of classification.

[0052] As a further improvement, in other embodiments, the method may further include: S7, calculating the spatial position and distribution density of cracks, depressions and bulges based on the final classification results, and using a heat map generation algorithm to draw a runway surface damage distribution map to obtain a damage distribution map.

[0053] The identification data of cracks, dents, and bulges are obtained from the classification results to determine the spatial coordinates of each type of damage. Based on the spatial coordinates, the grid division algorithm is used to calculate the distribution density of cracks, dents, and bulges on the runway surface to obtain density distribution data. If there are outliers in the density distribution data, the outliers are processed by the mean filtering algorithm to obtain smooth density distribution data. A heat map generation algorithm is used to generate a heat map of the runway surface damage based on the smooth density distribution data to obtain heat map data. The damage distribution characteristics are extracted from the heat map data to determine the damage concentration area. Based on the damage concentration area, visual damage distribution data is generated to obtain a visualization result of the runway surface damage distribution. If the damage concentration area in the visualization result exceeds the preset threshold, the damage area is grouped by the clustering algorithm to obtain damage grouping data.

[0054] Specifically, first, the runway surface is scanned using an image processing algorithm to extract the features of cracks, depressions, and bulges.

[0055] For example, the Canny edge detection algorithm is used to identify cracks, and the threshold is set to 0.5. The detected crack length is 2.3 meters and the width is 0.1 meters. For depressions, a deep convolutional neural network (DCNN) is used for identification, and the depth threshold is set to 0.2 meters. The detected depression area is 1.5 square meters. The bulge is identified by the morphological processing algorithm, and the height threshold is set to 0.15 meters. The detected bulge area is 0.8 square meters. Then, based on these feature data, the spatial position and distribution density of each damage type are calculated.

[0056] For example, the distribution density of cracks on the east side of the runway is 0.3 per square meter, the distribution density of depressions in the middle of the runway is 0.2 per square meter, and the distribution density of bulges on the west side of the runway is 0.1 per square meter. Then, a heat map generation algorithm is used to draw a runway surface damage distribution map. The Gaussian kernel density estimation (KDE) algorithm is used to set the bandwidth to 0.5 to generate a heat map. In the heat map, the crack area is displayed in red, the depression area is displayed in yellow, and the bulge area is displayed in blue. The depth of color indicates the high and low density of damage. Finally, the heat map data is converted into visual data for further analysis and decision-making.

[0057] For example, the generated heat map data contains the damage type and density value of each pixel, which facilitates subsequent runway maintenance and repair work.

[0058] As a further improvement, in other embodiments, the method may further include: S8, extracting high-density damage areas from the damage distribution map, using a clustering algorithm to determine priority maintenance areas where cracks are dense or depressions are concentrated, and obtaining a maintenance priority list.

[0059] An initial data set is obtained from the damage distribution data, and the noise is removed by using a preprocessing technique to obtain a first damage data set. A damage distribution map is generated by using data visualization technology through the first damage data set to determine the damage distribution map. Features are extracted from the damage distribution map, and a clustering algorithm is used to identify high-density areas to obtain crack-dense areas and dent-dense areas. If the density of a crack-dense area is greater than a preset threshold, it is marked as a first priority maintenance area; if the density of a dent-dense area is greater than a preset threshold, it is marked as a second priority maintenance area to obtain a set of priority maintenance areas. According to the set of priority maintenance areas, a sorting algorithm is used to arrange them in descending order of density value to generate a maintenance priority sequence. The area identifier is extracted from the maintenance priority sequence to generate a maintenance priority list to obtain a final maintenance priority list. For the final maintenance priority list, the data storage technology is used to save it to the database to obtain the storage path.

[0060] Specifically, in the damage distribution map, the damaged area is first binarized using image processing technology, the damaged area is marked as 1, and the non-damaged area is marked as 0, forming a damage distribution matrix.

[0061] For example, the Otsu algorithm is used to automatically determine the binarization threshold to separate the damaged area from the background. Next, a density clustering algorithm (such as the DBSCAN algorithm) is used to cluster the damaged area, setting the neighborhood radius ε to 5 pixels and the minimum number of samples MinPts to 10 to identify high-density damaged areas. By calculating the center point coordinates and damage density of each cluster, the areas with dense cracks or concentrated depressions are determined.

[0062] For example, the coordinates of a cluster center are (120, 80) and the damage density is 0.85, indicating that the damage in this area is relatively concentrated. Then, the maintenance priority is calculated based on the damage density and the area. The weighted scoring method is used to set the damage density weight to 0.6 and the area weight to 0.4 to calculate the priority score of each area.

[0063] For example, if the damage density of a certain area is 0.85 and the area is 200 pixels, its priority score is 0.85×0.6+200×0.4=86.3. Finally, all areas are sorted from high to low according to the priority score to generate a maintenance priority list for reference by maintenance personnel.

[0064] For example, the priority list shows that area A has the highest score of 92.5, followed by area B at 88.7, and so on, ensuring that high-density damage areas are maintained first to improve maintenance efficiency.

[0065] As a further improvement, in other embodiments, the method may further include: S9, for the maintenance priority list, generating an optimized flight path for the next inspection of the drone through a path planning algorithm, reducing the frequency of repeated scanning of crack-dense areas, and obtaining new flight path data.

[0066] The detection area allocation data is obtained from the maintenance priority list, and the initial flight path data is generated through the path planning algorithm. If the initial flight path data covers the crack-dense area, the scanning frequency adjustment module is used to reduce the repeated scanning frequency to obtain the adjusted flight path data. The A* algorithm is used to optimize the path for the adjusted flight path data to obtain the optimized flight path data. The drone scheduling module is used to allocate drone detection tasks according to the optimized flight path data to generate drone task data. If there is a path conflict in the drone task data, the conflict detection module is used to reallocate the detection area to obtain the updated task data. Based on the updated task data, new flight path data is generated. Through the data generation module, the new flight path data is converted into path instructions executable by the drone to obtain the final path instruction data.

[0067] Specifically, in the maintenance priority list, the crack images collected by the drone are first analyzed using an image recognition algorithm to identify the density and distribution of the cracks.

[0068] For example, a convolutional neural network (CNN) is used to process the image and calculate the crack density value of each area. If the crack density in one area is 0.8 and in another area is 0.3, the area with higher crack density will be marked as a high priority. Next, based on these priority data, the A* path planning algorithm is used to generate the optimized flight path of the drone. The A* algorithm selects the optimal path by evaluating the cost function of each node. Assuming that the current position of the drone is (0,0) and the target position is (10,10), the A* algorithm will calculate a path that avoids high-density crack areas, such as (0,0)→(2,2)→(4,4)→(6,6)→(8,8)→(10,10). In order to further reduce the frequency of repeated scanning in crack-dense areas, a dynamic weight adjustment mechanism is introduced to adjust the path weight in real time according to the changes in crack density.

[0069] For example, when the crack density in a certain area decreases from 0.8 to 0.5, the system will automatically reduce the scanning frequency of the area and allocate more resources to other high-density areas. Ultimately, through the above algorithm and analysis process, new flight path data is generated to ensure that the drone can complete the inspection task efficiently and accurately, while reducing repeated scanning of high-density crack areas.

[0070] S10, controlling the flight of the UAV through the new flight path data, using sensor calibration to reduce deviations caused by vibration or temperature difference, and transmitting the collected high-resolution images to a ground processing unit in combination with real-time data transmission to obtain an updated first image set.

[0071] Vibration and temperature difference deviation data are obtained through sensor calibration, and the preset calibration model is used to correct the sensor output to obtain calibrated sensor data. According to the calibrated sensor data and combined with the new flight path data, the path planning algorithm is used to generate the UAV control instructions to determine the flight trajectory of the UAV. If the UAV flight trajectory deviates from the preset path, the control instructions are adjusted through the real-time feedback mechanism to obtain a stable flight state. High-resolution images are collected through the onboard camera, and the collected image data are processed by the image compression algorithm to obtain compressed image data. The compressed image data is sent to the ground processing unit through the real-time transmission protocol to obtain the transmitted image data. The transmitted image data is decompressed and denoised in the ground processing unit to obtain the first image set. For the first image set, the image segmentation algorithm is used to extract the target area to obtain the processed target image data.

[0072] Specifically, when the new flight path data is used to control the flight of the drone, a path planning algorithm based on B-spline curves is used to generate a smooth trajectory, where the control point spacing is set to 5 meters and the order k=3, ensuring that the turning radius of the drone is not less than 2 meters to avoid excessive centrifugal force. The sensor calibration phase uses Kalman filtering to fuse IMU and GPS data. For the attitude deviation of ±0.5°, real-time correction is performed through the state transfer matrix Q=diag[0.01,0.01,0.01] and the observation noise matrix R=diag[0.1,0.1]. In order to eliminate the lens distortion caused by temperature difference, a polynomial fitting model is used to compensate for the -20℃ to 50℃ data collected by the infrared temperature sensor, and a distortion correction equation of δ=0.03T² - 0.12T + 0.15 is established. The real-time data transmission uses H.265 encoding to compress 4K / 60fps images, and the bit rate is controlled at 20Mbps. The LDPC (2048,1024) forward error correction coding ensures that the bit error rate is less than 1e-6. After receiving the data, the ground processing unit first uses the SIFT feature matching algorithm to align the continuous frames. When the number of feature points is greater than 500 pairs, the Bundle Adjustment optimization is performed, and the reprojection error converges to within 0.3 pixels. Finally, an orthogonal correction image with a resolution of 2048×2048 is output. In terms of vibration suppression, at a sampling frequency of 400Hz, the main vibration frequency band is identified as 8-12Hz through fast Fourier transform, and an IIR bandstop filter is used to set -40dB attenuation at the center frequency of 10Hz. All processing processes are completed in the embedded system, and the end-to-end delay from data acquisition to image output is controlled within 120ms.

[0073] It should be understood by those skilled in the art that the above embodiments are only for the purpose of clearly illustrating the present invention, and are not intended to limit the scope of the present invention. For those skilled in the art, other changes or modifications may be made based on the above disclosure, and these changes or modifications are still within the scope of the present invention.

Claims

1. A method for identifying sports track pavement damage, characterized in that: The method comprises: Establishing construction defect templates and using worn templates; Controlling the UAV to fly along the runway geometry through a preset flight path, using automatic exposure adjustment and shutter speed optimization to adapt to cloudy or bright light conditions, and collecting high-resolution runway surface images to obtain a first image set; For the first image set, bilateral filtering is used to smooth light interference noise and retain edge details, and pixel value normalization is used to correct the influence of strong light reflection or shadow to obtain the second image set; Defect features are extracted from the second image set, and multi-scale feature analysis is performed using a convolutional neural network to calculate the crack length, the concave area, and the bulge height, generate a high-dimensional feature vector containing the defect shape and texture, and obtain a defect feature set; If the similarity between the feature vector in the defect feature set and the construction defect template is higher than the first threshold value T1, it is judged as a construction defect and the first classification result is obtained; otherwise, the subsequent judgment is performed; If the similarity between the feature vector in the defect feature set and the wear template is higher than the second threshold T2, it is judged as wear and tear and a second classification result is obtained; otherwise, it is marked as an unclassified defect and a third classification result is obtained.

2. The method for identifying sports track pavement damage according to claim 1, characterized in that: in, The first threshold T1 is set to 0.85-0.9, and the second threshold T2 is set to 0.75-0.

8.

3. The method for identifying sports track pavement damage according to claim 2, characterized in that: Further including: For the first and second classification results, support vector machine is used to map high-dimensional feature space through kernel function, optimize the classification boundaries of construction defects and wear and tear, and obtain the final classification results.

4. The method for identifying sports track pavement damage according to claim 3, characterized in that: Further including: According to the final classification results, the spatial position and distribution density of cracks, depressions and bulges are calculated, and the heat map generation algorithm is used to draw the runway surface damage distribution map.

5. The method for identifying sports track pavement damage according to claim 4, characterized in that: Further including: Extract high-density damage areas from the damage distribution map, use clustering algorithm to determine the priority maintenance areas with dense cracks or concentrated depressions, and obtain a maintenance priority list; Based on the maintenance priority list, the optimized flight path for the next inspection of the UAV is generated through the path planning algorithm, and the new flight path data is obtained to reduce the frequency of repeated scanning in the crack-dense area; The new flight path data is used to control the flight of the UAV, and the collected high-resolution images are transmitted to the ground processing unit in combination with real-time data transmission to obtain an updated first image set.

6. The method for identifying sports track pavement damage according to claim 1, characterized in that: The method of obtaining the second image set includes: Obtaining initial pixel data from the first image set, using bilateral filtering to process light interference noise, retaining edge details, and obtaining an intermediate image set; For the intermediate image set, if the pixel value exceeds the preset strong light threshold, the strong light reflection is corrected by normalizing the pixel value to obtain a corrected image set; For the correction image set, if a shadow area is detected, the shadow area is adjusted by using local contrast enhancement to obtain an enhanced image set; Extract edge features from the enhanced image set, use edge detection algorithm to determine the integrity of edge details, and obtain edge feature sets; According to the edge feature set, the mean filter is used to smooth the noise in the non-edge area to obtain a smoothed image set; For the smoothed image set, the overall brightness distribution is optimized by histogram equalization to obtain the second image set.

7. The method for identifying sports track pavement damage according to claim 1, characterized in that: The method for obtaining the defect feature set includes: Acquire a defect image from the second image set, remove noise using image preprocessing technology, and obtain a clear defect image; The multi-scale feature extraction of the clear defect image is performed through the convolutional neural network to generate a multi-scale feature map; For the multi-scale feature map, the crack length, depression area and bulge height are calculated to obtain the defect geometric parameters; If the defect geometric parameters exceed the preset threshold, the texture analysis algorithm is used to extract the defect texture features and generate a texture feature set; According to the defect geometric parameters and the texture feature set, a high-dimensional feature vector including the defect shape and texture is constructed to obtain a feature vector set; Classify the feature vector set through clustering algorithm to determine the defect feature set; If the defect feature set matches the preset defect template, the feature fusion technology is used to generate the final defect feature set.

8. The method for identifying sports track pavement damage according to claim 4, Features: Among them, according to the final classification results, the spatial position and distribution density of cracks, depressions and bulges are calculated, and the method of using the heat map generation algorithm to draw the runway surface damage distribution map includes: Obtain identification data of cracks, dents and bulges from the classification results, and determine the spatial coordinates of various types of damage; According to the spatial coordinates, the distribution density of cracks, depressions and bulges on the runway surface is calculated using a grid division algorithm to obtain density distribution data; If there are outliers in the density distribution data, the outliers are processed by the mean filtering algorithm to obtain smooth density distribution data; A heat map generation algorithm is used to generate a heat map of the runway surface damage based on the smoothed density distribution data to obtain heat map data; Extract damage distribution characteristics from heat map data and determine the damage concentration area; Generate visual damage distribution data based on the damage concentration area and obtain the visualization results of the runway surface damage distribution; If the damage concentration area in the visualization result exceeds the preset threshold, the damage area is grouped by a clustering algorithm to obtain damage grouping data.