Asphalt pavement looseness quantitative evaluation method based on vehicle-mounted image acquisition

Abstract

The invention discloses an asphalt pavement looseness quantitative evaluation method based on vehicle-mounted image acquisition. The method comprises the following steps: acquiring a pavement image; carrying out distortion correction processing on the pavement image, and carrying out region frame selection on the processed image to obtain a target pavement image region; carrying out segmentation processing on the target pavement image region to obtain a pavement aggregate region; and taking the ratio of the area of the pavement aggregate region to the area of the target pavement image region as a loosening quantitative evaluation parameter to obtain the loosening grade of the pavement. According to the invention, quantitative evaluation can be carried out on the pavement looseness, and the evaluation efficiency and accuracy of the pavement looseness are improved.

Description

Technical Field

[0001] The invention relates to the field of road engineering, and in particular to a method for quantitatively evaluating looseness of an asphalt pavement based on vehicle-mounted image acquisition. Background Art

[0002] Asphalt pavement loosening is a common form of asphalt pavement damage, which directly affects the performance and life of the road. Its occurrence is usually related to a variety of factors, including raw material quality, construction technology, traffic load and environmental factors. In particular, the quality of asphalt will directly affect the bonding performance between asphalt and aggregate, and then affect the durability of the pavement. At present, in engineering practice, the evaluation of pavement loosening mainly relies on visual inspection and empirical judgment, combined with a lightweight falling weight instrument (such as FWD, Falling Weight Deflectometer) to measure the rebound modulus of the pavement structure to indirectly evaluate the looseness. In addition, some projects also use geological radar (GPR) and material sampling tests (such as pull-out tests) to perform non-destructive testing and laboratory analysis on the integrity of the internal structure of the pavement.

[0003] Visual inspection and empirical judgment are the most traditional and commonly used evaluation methods. Experienced engineers usually judge the looseness of the pavement surface through on-site visual inspection, such as surface particle shedding, increased cracks, and sanding of the pavement. Although simple, easy to operate and low-cost, this method relies on personal experience, is highly subjective, and is difficult to quantify, resulting in inconsistent evaluation results. FWD simulates vehicle loads and measures the modulus of resilience of the pavement under impact loads, thereby indirectly evaluating the looseness of the pavement structure. The test results of FWD can reflect the overall stiffness and bearing capacity of the pavement structure and can provide a certain degree of quantitative data, but because it mainly focuses on the overall structural performance, it cannot accurately locate and quantitatively evaluate the specific situation of surface looseness. Material sampling and laboratory testing directly test the bond strength and shear strength of the material by sampling pavement materials and conducting laboratory pull-out tests, shear tests, etc. to evaluate the root cause of looseness. This method is often used for in-depth analysis of severely loose areas. Laboratory testing can provide very accurate quantitative data, but the sampling and testing process is complicated and expensive, and the test results only represent the local situation of the sampling point, which is difficult to reflect the overall pavement status.

[0004] In summary, the existing pavement looseness assessment lacks quantitative assessment methods, and its pavement looseness mostly relies on empirical inspections and indirect mechanical tests. These methods have the problems of strong subjectivity and low accuracy. Therefore, in order to solve the above problems, a new asphalt pavement looseness quantitative assessment method is needed, which can quantitatively assess the looseness of the pavement and improve the efficiency and accuracy of the pavement looseness assessment. Summary of the invention

[0005] In view of this, the purpose of the present invention is to overcome the defects in the prior art and provide a quantitative assessment method for asphalt pavement looseness based on vehicle-mounted image acquisition, which can quantitatively evaluate the looseness of the pavement and improve the efficiency and accuracy of the assessment of the looseness of the pavement.

[0006] The method for quantitatively evaluating looseness of asphalt pavement based on vehicle-mounted image acquisition of the present invention comprises:

[0007] Collect road surface images; perform distortion correction processing on the road surface images, select regions on the processed images, and obtain target road surface image regions;

[0008] Segment the target road surface image area to obtain the road surface aggregate area;

[0009] The ratio of the pavement aggregate area to the target pavement image area is used as a loose quantitative evaluation parameter to obtain the looseness grade of the pavement.

[0010] Further, collecting road surface images specifically includes:

[0011] Select an industrial camera with resolution, frame rate, and optical characteristics that meet your target requirements;

[0012] The industrial camera is fixedly mounted at the center of the vehicle chassis so that its line of sight is vertically downward to obtain a top view of the road surface; wherein the camera is mounted at a height of about 1.5 to 2 meters from the ground and is provided with shockproof protection;

[0013] Reasonably set the camera's exposure time, set or automatically adjust the camera's white balance parameters according to the on-site lighting conditions, and set the camera's shutter speed according to the vehicle's driving speed;

[0014] The camera is set to continuous shooting mode, and the shooting interval is adjusted according to the vehicle speed and image overlap requirements;

[0015] The captured image is transmitted via a high-speed data line. If blurry or improperly exposed images are detected, the camera parameters are automatically adjusted or the image is recaptured.

[0016] Further, the road surface image is subjected to distortion correction processing, specifically including:

[0017] Use edge detection to identify the edges in the image, and apply the line detection algorithm to detect the position of the lane lines in the image;

[0018] Perform curve fitting on the detected lane edge points;

[0019] The fitted curve is mapped back into straight line form by computing the inverse transform:

[0020] Use perspective transformation to construct a perspective transformation matrix, which is used to convert points in the image into corrected points. By adjusting the parameters of the perspective transformation matrix, the distortion in the image is corrected so that the lane lines appear as straight lines.

[0021] Furthermore, the target road surface image area is segmented using an image semantic segmentation model;

[0022] The image semantic segmentation model uses the DeepLabv3+ model as the baseline, including: dilated convolution in the convolutional neural network, dilated spatial pyramid pooling module, and pre-trained Xception deep convolutional neural network;

[0023] The dilated convolution expands the receptive field of the convolution kernel without increasing the number of parameters and the amount of calculation by inserting holes between the convolution kernel elements.

[0024] The atrous spatial pyramid pooling module integrates features from different scales by applying multiple convolution operations with different atrous rates in parallel, generating a feature map with rich contextual information, and further enhancing the multi-scale feature extraction capability of the model.

[0025] The Xception deep convolutional neural network includes convolutional layers and maximum pooling layers, an intermediate stream and an exit stream; wherein the input image first passes through a standard convolutional layer and a maximum pooling layer to preliminarily extract low-level features and reduce the size of the feature map; the intermediate stream is composed of multiple depth-separable convolutional blocks, each of which contains depth-wise convolution and point-by-point convolution, and the output feature map gradually shrinks in space but increases in channel dimension, and the intermediate stream is responsible for extracting high-level semantic features; the exit stream is used to further process the features extracted by the intermediate stream, including several layers of depth-separable convolution and global average pooling layers, and the output of the exit stream is used for final classification or other tasks. In DeepLabv3+, it passes the features to the subsequent dilated spatial pyramid pooling module and decoder part.

[0026] Furthermore, the image semantic segmentation model selects a cross entropy loss function for training; the cross entropy loss function is used to measure the difference between the category probability distribution output by the model and the true label;

[0027] In the semantic segmentation task, the input image is divided into multiple pixels, and the goal of the model is to predict a class label for each pixel; for each pixel, the cross entropy loss calculates the distance between the probability distribution predicted by the model and the true distribution.

[0028] Furthermore, the ratio of the pavement aggregate area to the target pavement image area is used as a loose quantitative evaluation parameter to obtain the looseness level of the pavement, including:

[0029] Construct loose quantitative evaluation indicators:

[0030] RA = 1-SA / SS;

[0031] Among them, RA is the proportion of loose area; SA is the area of ​​pavement aggregate area; SS is the area of ​​target pavement image area;

[0032] A loose evaluation grade table is set, and the target grade corresponding to the loose quantitative evaluation indicator is found from the loose evaluation grade table.

[0033] Furthermore, the pavement aggregate area is determined according to the following method:

[0034] Select a calibration plate with known dimensions and set marking points on it;

[0035] Place the calibration plate on the ground within the field of view of the vehicle camera;

[0036] Using image processing algorithms to detect the positions of the marking points of the calibration plate in the image;

[0037] According to the position information, the pixel distance of the calibration plate in the image is calculated to obtain the calibration factor of the pixel size: Among them, d i is the pixel distance between two adjacent markers, n is the number of pixel distances, D actual It is the actual physical length between two adjacent marking points;

[0038] Calculate the pixel area of ​​the pavement aggregate area in the image and obtain the pavement aggregate area SA:

[0039]

[0040] The beneficial effects of the present invention are as follows: the present invention discloses a method for quantitatively evaluating the looseness of asphalt pavement based on vehicle-mounted image acquisition, which uses a vehicle-mounted industrial camera to acquire high-definition images of the pavement in real time, determines and segments the pavement area, uses a distortion correction method to obtain a bird's-eye view image of the pavement, and uses a deep learning semantic segmentation algorithm to achieve the segmentation of pavement aggregates, calibrate the actual size of the pavement image, calculate the area ratio of the identified pavement aggregate part to the overall photographed pavement, determine the looseness quantitative evaluation index, and realize an automated and refined evaluation of the looseness of the pavement. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0042] Figure 1 It is a schematic diagram of the process flow of the method for quantitatively evaluating road looseness of the present invention. DETAILED DESCRIPTION

[0043] The present invention is further described below in conjunction with the accompanying drawings, as shown in the drawings:

[0044] This embodiment discloses a method for quantitatively evaluating looseness of an asphalt pavement based on vehicle-mounted image acquisition, comprising the following steps:

[0045] Collect road surface images; perform distortion correction processing on the road surface images, select regions on the processed images, and obtain target road surface image regions;

[0046] Segment the target road surface image area to obtain the road surface aggregate area;

[0047] The ratio of the pavement aggregate area to the target pavement image area is used as a loose quantitative evaluation parameter to obtain the looseness grade of the pavement.

[0048] In this embodiment, the present invention uses a high-resolution industrial camera to ensure that road surface details such as the distribution and shape of aggregate particles can be captured. The selection of the camera is mainly based on the following considerations:

[0049] Resolution: The camera should have a resolution of at least 20 megapixels to ensure that the clarity of the captured image can meet the needs of the subsequent deep learning algorithm for aggregate segmentation. Frame rate: The camera should have a high frame rate (at least 30 frames per second) to continuously capture images while the vehicle is driving to avoid image blur or distortion caused by excessive speed. Optical properties: Select a camera with a low distortion lens to reduce image distortion caused by lens distortion and ensure that the road geometry in the image is as close to the actual situation as possible. Camera installation: The camera is fixedly installed in the center of the vehicle chassis, ensuring that its line of sight is vertically downward to obtain a top view of the road surface. Installation height: The camera is installed at a height of about 1.5 to 2 meters from the ground, so that it can capture a large enough field of view and ensure that the image resolution is sufficient for detailed analysis. Shockproof design: In order to cope with the vibration of the vehicle during driving, the camera mounting system needs to be designed with a special shockproof structure, such as using a shock-absorbing bracket or installing a shock-proof pad to reduce image blur caused by uneven road surface.

[0050] The image acquisition parameters are set as follows:

[0051] Exposure time: Set the exposure time appropriately to avoid overexposure or underexposure of the image under different lighting conditions (such as direct sunlight or shadow coverage). Generally speaking, you can select the automatic exposure mode or preset the appropriate exposure value under fixed environment. White balance: Set or automatically adjust the camera's white balance parameters according to the lighting conditions on site to ensure that the color of the image is restored realistically, especially the color and texture of the asphalt road surface can be accurately presented. Shutter speed: When the vehicle speed does not exceed 60 km / h, the shutter speed is set to 1 / 1000 second or faster to ensure a clear image and avoid motion blur.

[0052] During data acquisition, the camera is set to continuous shooting mode to ensure that road images can be continuously obtained while the vehicle is in motion. The shooting interval is adjusted according to the vehicle speed and image overlap requirements, and is usually set to shoot 5 to 10 images per second to ensure that there is appropriate overlap between images to avoid missing road areas. The image data is transmitted to the on-board computer system via a high-speed data line or directly stored in the camera's built-in memory. The on-board computer must have sufficient storage capacity and data processing capabilities to cope with the real-time storage and processing requirements of large-capacity image files. The on-board computer system monitors the image acquisition process in real time to ensure that the image quality meets the requirements. If blurry or improper exposure is detected, the system should issue an alarm and automatically adjust the camera parameters, or re-acquire the image.

[0053] In addition, lighting conditions can change significantly during different time periods and weather conditions. To address this challenge, optical filters can be used during acquisition, or brightness correction can be performed through post-image processing. In addition, you can consider installing lighting equipment on the front of the vehicle to ensure consistent lighting conditions. Changes in vehicle speed may lead to inconsistent image spacing, affecting subsequent image stitching and analysis. To this end, the camera's shooting frequency can be adjusted in real time in combination with the vehicle's GPS data or speed sensor data to ensure consistent image acquisition. Dust, rain, snow, etc. on the road may affect image quality. In this case, image preprocessing techniques such as denoising and defogging can be used to improve image clarity and usability.

[0054] In this embodiment, in the road image, the lane line is usually considered to be a straight line segment, especially on a straight road segment. Using this prior knowledge, the image distortion correction can be effectively performed. The distortion mainly comes from the geometric distortion of the camera lens, especially when the on-board camera uses a wide-angle lens. This distortion will cause the straight line to appear curved in the image. Assuming ideal conditions, the lane line should appear in a straight line shape in the image.

[0055] The present invention detects and fits the lane lines in the image as curves, and corrects them into straight lines through inverse calculation. Assume that a lane line in the image can be represented as a curve y=f(x) on a two-dimensional plane, and ideally this line should be y=kx+b (where k and b are the slope and intercept of the line, respectively).

[0056] In the specific distortion correction process, first, use image processing techniques such as Canny edge detection to identify the edges in the image, and then apply Hough transform or other line detection algorithms to detect the lane line positions in the image. For curve fitting of the detected lane line edge points, a polynomial fitting method, such as a quadratic or cubic polynomial, can be used to obtain the curve equation y=f(x).

[0057] By calculating the inverse transformation, the fitted curve is mapped back to the ideal straight line form. Specifically, find the transformation matrix T in the image so that the mapped curve becomes y=kx+b. This goal can be achieved using perspective transformation. The general form of the perspective transformation matrix M is:

[0058]

[0059] This matrix is ​​used to transform a point (x, y) in the image to a rectified point (x', y'):

[0060]

[0061] By adjusting the parameters of the perspective transformation matrix M, the distortion in the image can be corrected so that the lane lines appear as straight lines.

[0062] In this embodiment, before manual frame selection is performed, the image should be distorted and corrected to ensure the front view effect of the viewing angle. In order to facilitate the subsequent frame selection operation, the image needs to be pre-processed in a basic manner, such as grayscale, contrast enhancement, etc.

[0063] Use interactive image processing software, such as the ROI (Region of Interest) tool in Matlab and OpenCV, to manually select the road area to be analyzed in the image. Users can manually outline rectangular or polygonal areas based on the road characteristics to be analyzed, ensuring that the selected area contains complete lanes and loose detection target areas. The selection process should try to avoid including interference objects other than lane lines, such as road signs or curbs.

[0064] In this embodiment, the target pavement image area is segmented using an image semantic segmentation model to obtain two parts: aggregate and background. The pavement aggregate area is accurately identified, the aggregate part is retained, the background part is removed, and the segmentation effect and robustness of the model under different pavement conditions are evaluated.

[0065] The image semantic segmentation model uses the DeepLabv3+ model as the baseline, including: dilated convolution in the convolutional neural network, dilated spatial pyramid pooling module, and pre-trained Xception deep convolutional neural network;

[0066] The design goal of DeepLabv3+ is to improve the segmentation performance of the model in complex scenes by introducing multi-scale context information, while restoring high-resolution feature maps through the decoder structure to ensure the precision of the segmentation results. DeepLabv3+ performs well in many benchmarks, especially for tasks that require processing complex backgrounds and multi-scale objects.

[0067] The core architecture of DeepLabv3+ is an encoder-decoder structure. The encoder is responsible for gradually reducing the spatial resolution of the image while extracting more and more abstract features. The decoder is responsible for restoring these abstract features to high resolution to produce fine-grained segmentation results.

[0068] The encoder part mainly consists of a deep convolutional neural network (usually a pre-trained ResNet or Xception) and an atrous convolution layer. The pre-trained deep convolutional neural network is used to extract preliminary image features, while the atrous convolution layer expands the receptive field by adjusting the expansion rate of the convolution kernel (called the atrous rate), capturing more contextual information without increasing the amount of computation. Atrous convolution is a special convolution operation that expands the receptive field of the convolution kernel by introducing "holes" between the convolution kernels, that is, inserting several zeros between the kernel elements. This method can process a wider range of contextual information without significantly increasing the computational cost.

[0069] The decoder part is designed to gradually restore the spatial details lost in the encoder. DeepLabv3+ uses a simple but effective decoder structure to fuse the high-resolution low-level features of the encoder with the low-resolution high-level features to generate high-quality segmentation results. The decoder first upsamples the low-resolution feature map output by the encoder (usually 4 times), and then concatenates it pixel by pixel with the high-resolution feature map extracted from the early layers of the encoder. The concatenated feature map is processed through a series of convolutional layers to finally generate a predicted segmentation map with the same resolution as the original image.

[0070] DeepLabv3+ overcomes the problem of blurred segmentation maps or unclear boundaries that may be caused by relying solely on high-level features by fusing low-level high-resolution features with deep-level low-resolution features. In terms of feature fusion, by splicing feature maps of different levels, the decoder can make full use of the detail information and context information extracted by each layer in the encoder. After feature fusion, the clarity and accuracy of the segmentation map can be further improved by processing the convolution layer. In terms of upsampling strategies, the upsampling operation in the decoder usually uses simple and efficient methods such as bilinear interpolation to enlarge the low-resolution feature map to the original image resolution. Although this method is relatively simple, combined with the splicing operation of high-resolution features, it can significantly improve the refinement of the segmentation map.

[0071] The image semantic segmentation model uses the feature extraction module to extract multi-scale and multi-level features from the input image and use these features to generate accurate segmentation maps. The image semantic segmentation model combines the dilated convolution and dilated spatial pyramid pooling (ASPP) modules in the convolutional neural network (CNN) with the pre-trained Xception deep convolutional neural network through DeepLabv3+, which together form its powerful feature extraction capabilities.

[0072] The structure of the Xception network consists of a series of depth-separable convolutional layers, which extract complex features of the image layer by layer while maintaining efficient computation. The Xception network usually consists of the following parts:

[0073] Initial convolution layer and maximum pooling layer: The input image first passes through a standard convolution layer and a maximum pooling layer to initially extract low-level features and reduce the size of the feature map, laying the foundation for subsequent depthwise separable convolution.

[0074] Middle Flow: The middle flow part is the main feature extraction module of the Xception network, which consists of multiple depth-separable convolution blocks. Each convolution block contains depth-wise convolution and point-wise convolution, and the output feature map gradually shrinks in space, but increases in channel dimension. This part is responsible for extracting high-level semantic features.

[0075] Exit Flow: The exit flow part further processes the features extracted by the intermediate flow, usually including several layers of depth-separable convolution and global average pooling layers. The output of the exit flow is used for the final classification or other tasks. In DeepLabv3+, it passes the features to the subsequent ASPP module and decoder part.

[0076] The dilated convolution expands the receptive field of the convolution kernel without increasing the number of parameters and computation by inserting holes between the convolution kernel elements (i.e. skipping some pixels). This feature gives the dilated convolution a significant advantage in extracting multi-scale features. In terms of expanding the receptive field: the dilated convolution expands the pixel range covered by the convolution operation by increasing the dilation rate of the convolution kernel, so that the network can perceive a wider range of contextual information with the same amount of computation. This is particularly important for processing images with multi-scale structures. For example, in segmentation tasks, the dilated convolution can capture both the global background and local details. In terms of maintaining the resolution of the feature map, unlike traditional downsampling operations, the dilated convolution does not reduce the resolution of the feature map while expanding the receptive field. This means that the network can process more complex semantic information while retaining image details, which is especially suitable for tasks such as small object segmentation and edge detection.

[0077] The Atrous Spatial Pyramid Pooling (ASPP) module is also an important part, which aims to further enhance the multi-scale feature extraction capability of the model. The ASPP module integrates features from different scales by applying multiple convolution operations with different dilation rates in parallel, thereby generating a feature map with rich contextual information. In terms of multi-scale atrous convolution, the ASPP module contains multiple parallel atrous convolutions, each with a different dilation rate (such as 6, 12, 18). This design allows the network to process multi-scale information from local details to global background simultaneously. For example, a convolution kernel with a dilation rate of 6 may be more suitable for capturing subtle texture features, while a convolution kernel with a dilation rate of 18 helps capture the global structure of the image. In terms of global average pooling, in addition to the atrous convolution, the ASPP module also includes a global average pooling layer for extracting global features. Global average pooling captures the overall semantic information of the image by compressing the entire feature map into a global feature vector by performing a pooling operation. This feature, when spliced ​​with other multi-scale features, can further enrich the contextual understanding capability of the model. In terms of feature fusion, the multi-scale features output by the ASPP module are channel compressed through 1x1 convolution and then concatenated with the output feature map of the basic convolutional network. The resulting feature map contains rich global context information while retaining the image details, providing high-quality input for the subsequent decoder part.

[0078] In this embodiment, the loss function plays a crucial role in the training process of the model. For this reason, the cross-entropy loss function (Cross-Entropy Loss) is selected for training. This loss function can measure the difference between the class probability distribution output by the model and the true label.

[0079] In the semantic segmentation task, the input image is divided into multiple pixels, and the goal of the model is to predict a category label for each pixel. Assuming there are C categories, the output of the model is a probability vector p = [p1, p2, ..., p c ], where p i represents the probability that a pixel belongs to the i-th class. The true label is represented by a one-hot vector y=[y1, y2, ..., y C ] means that only the position corresponding to the true category is 1, and the rest are 0.

[0080] The cross entropy loss function can be expressed as:

[0081]

[0082] For each pixel, the cross entropy loss calculates the distance between the probability distribution predicted by the model and the true distribution. The final loss is the average loss over all pixels:

[0083]

[0084] Where N is the number of pixels in the image, L n is the cross entropy loss of the n-th pixel.

[0085] In this embodiment, the ratio of the road aggregate area to the target road image area is used as a loose quantitative evaluation parameter to obtain the looseness level of the road surface, which specifically includes:

[0086] Construct loose quantitative evaluation indicators:

[0087] RA = 1-SA / SS;

[0088] Among them, RA is the proportion of loose area; SA is the area of ​​pavement aggregate; SS is the area of ​​target pavement image area;

[0089] The area of ​​the pavement aggregate area is determined according to the following method: In order to convert the pixel information segmented from the image into the actual physical size (such as centimeters or meters), the image size calibration is required. A calibration plate of known size is placed in the shooting scene, and then image processing technology is used to determine the actual physical size corresponding to each pixel in the image. It mainly includes four steps: calibration plate preparation, calibration point detection, actual size calibration, and image size conversion.

[0090] Step 1: Prepare the calibration plate

[0091] Calibration plate selection: Select a calibration plate with known dimensions (e.g., a 1 meter x 1 meter square plate). The calibration plate should have obvious markings, usually cross lines or dots, so that it can be easily detected and identified in the image. Calibration plate placement: Place the calibration plate on the ground within the field of view of the vehicle camera, making sure that the calibration plate is level and completely within the camera's field of view. Keep the vehicle still during the shooting process to reduce the effect of motion blur.

[0092] Step 2: Calibration point detection

[0093] Calibration point detection: Use image processing algorithms (such as Hough transform or circle detection algorithm) to detect the positions of the four corner points or marking points of the calibration plate in the image. Assume that the pixel coordinates of the four corner points of the calibration plate in the image are: (x1, y1), (x2, y2), (x3, y3), (x4, y4).

[0094] Side length calculation: calculate the side length of the calibration plate in the image.

[0095] For example: Among them, d 12 is the pixel distance between two adjacent corner points of the calibration plate in the image.

[0096] Step 3: Actual size calibration

[0097] In order to improve the accuracy, the calibration factors of different sides of the calibration plate can be calculated and their average value can be taken as the final calibration factor:

[0098]

[0099] Where n is the number of sides of the calibration plate, d i is the pixel length of each edge in the image. The final calibration factor includes the case of single calibration, assuming that the actual side length of the calibration plate is D actual (e.g. 1 meter), then the pixel size calibration factor S in the image can be calculated pixel : This factor represents the actual physical length corresponding to each pixel.

[0100] Step 4: Image size conversion

[0101] Conversion formula: Once the calibration factor S is obtained pixel , which means that any measured pixel distance d in the image pixel Converted to actual physical distance D actual :

[0102] D actual =S pixel ×d pixel ;

[0103] Based on the above calibration method, the area calculation function in the image processing software (such as OpenCV, etc.) can be used to calculate the pixel area of ​​the pavement aggregate area in the image, and the image size conversion ratio S obtained by calibration can be used. pixel , convert the pixel area into the actual area and obtain the pavement aggregate area SA:

[0104]

[0105] A looseness evaluation grade table is set, and the target grade corresponding to the looseness quantitative evaluation index is found from the looseness evaluation grade table. The looseness evaluation grade table is shown in Table 1. The looseness quantitative evaluation index is the proportion of loose areas (looseness rate), and the looseness grades are divided according to the proportion of loose areas. The specific division standards can be fine-tuned according to actual maintenance needs.

[0106] Table 1

[0107] Loose area ratio Loose evaluation level (0，20％] No loose (20％,35％] Slightly loose (35％,50％] Moderately loose (50％,100％) Severe loose

[0108] This paper proposes an efficient and accurate quantitative assessment method for loose asphalt pavement by innovatively combining on-board image acquisition with the DeepLabv3+ deep learning algorithm. Compared with the traditional method that relies on manual inspection, this method can realize automatic detection of a large range of pavement areas without interfering with traffic. By accurately identifying and segmenting pavement aggregates and calculating the loose rate as a quantitative indicator, it not only improves the accuracy and efficiency of detection, but also provides a reliable scientific basis for road maintenance and management. This method has good scalability, is applicable to a variety of road types and environmental conditions, and significantly reduces maintenance costs.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A method for quantitatively evaluating looseness of asphalt pavement based on vehicle-mounted image acquisition, characterized in that: include: Collect road surface images; Perform distortion correction processing on the road surface image, select the area of ​​the processed image, and obtain the target road surface image area; Segment the target road surface image area to obtain the road surface aggregate area; The ratio of the pavement aggregate area to the target pavement image area is used as a loose quantitative evaluation parameter to obtain the looseness grade of the pavement.

2. The method for quantitatively evaluating looseness of asphalt pavement based on vehicle-mounted image acquisition according to claim 1 is characterized in that: Collect road images, including: Select an industrial camera with resolution, frame rate, and optical characteristics that meet your target requirements; The industrial camera is fixedly mounted at the center of the vehicle chassis so that its line of sight is vertically downward to obtain a top view of the road surface; wherein the camera is mounted at a height of about 1.5 to 2 meters from the ground and is provided with shockproof protection; Reasonably set the camera's exposure time, set or automatically adjust the camera's white balance parameters according to the on-site lighting conditions, and set the camera's shutter speed according to the vehicle's driving speed; The camera is set to continuous shooting mode, and the shooting interval is adjusted according to the vehicle speed and image overlap requirements; The captured image is transmitted via a high-speed data line. If blurry or improperly exposed images are detected, the camera parameters are automatically adjusted or the image is recaptured.

3. The method for quantitatively evaluating looseness of asphalt pavement based on vehicle-mounted image acquisition according to claim 1 is characterized in that: The road surface image is subjected to distortion correction processing, specifically including: Use edge detection to identify the edges in the image, and apply the line detection algorithm to detect the position of the lane lines in the image; Perform curve fitting on the detected lane edge points; The fitted curve is mapped back into straight line form by computing the inverse transform: Use perspective transformation to construct a perspective transformation matrix, which is used to convert points in the image into corrected points. By adjusting the parameters of the perspective transformation matrix, the distortion in the image is corrected so that the lane lines appear as straight lines.

4. The method for quantitatively evaluating looseness of asphalt pavement based on vehicle-mounted image acquisition according to claim 1 is characterized in that: Use the image semantic segmentation model to segment the target road image area; The image semantic segmentation model uses the DeepLabv3+ model as the baseline, including: dilated convolution in the convolutional neural network, dilated spatial pyramid pooling module, and pre-trained Xception deep convolutional neural network; The dilated convolution expands the receptive field of the convolution kernel without increasing the number of parameters and the amount of calculation by inserting holes between the convolution kernel elements. The atrous spatial pyramid pooling module integrates features from different scales by applying multiple convolution operations with different atrous rates in parallel, generating a feature map with rich contextual information, and further enhancing the multi-scale feature extraction capability of the model; The Xception deep convolutional neural network includes convolutional layers and maximum pooling layers, an intermediate stream and an exit stream; wherein the input image first passes through a standard convolutional layer and a maximum pooling layer to preliminarily extract low-level features and reduce the size of the feature map; the intermediate stream is composed of multiple depth-separable convolutional blocks, each of which contains depth-wise convolution and point-by-point convolution, and the output feature map gradually shrinks in space but increases in channel dimension, and the intermediate stream is responsible for extracting high-level semantic features; the exit stream is used to further process the features extracted by the intermediate stream, including several layers of depth-separable convolution and global average pooling layers, and the output of the exit stream is used for final classification or other tasks. In DeepLabv3+, it passes the features to the subsequent dilated spatial pyramid pooling module and decoder part.

5. The method for quantitatively evaluating looseness of asphalt pavement based on vehicle-mounted image acquisition according to claim 4 is characterized in that: The image semantic segmentation model selects a cross entropy loss function for training; the cross entropy loss function is used to measure the difference between the category probability distribution output by the model and the true label; In the semantic segmentation task, the input image is divided into multiple pixels, and the goal of the model is to predict a class label for each pixel; for each pixel, the cross entropy loss calculates the distance between the probability distribution predicted by the model and the true distribution.

6. The method for quantitatively evaluating looseness of asphalt pavement based on vehicle-mounted image acquisition according to claim 1 is characterized in that: The ratio of the pavement aggregate area to the target pavement image area is used as a loose quantitative evaluation parameter to obtain the looseness level of the pavement, including: Construct loose quantitative evaluation indicators: RA = 1-SA / SS; Among them, RA is the proportion of loose area; SA is the area of ​​pavement aggregate; SS is the area of ​​target pavement image area; A loose evaluation grade table is set, and the target grade corresponding to the loose quantitative evaluation indicator is found from the loose evaluation grade table.

7. The method for quantitatively evaluating looseness of asphalt pavement based on vehicle-mounted image acquisition according to claim 6 is characterized in that: Determine the pavement aggregate area as follows: Select a calibration plate with known dimensions and set marking points on it; Place the calibration plate on the ground within the field of view of the vehicle camera; Using image processing algorithms to detect the positions of the marking points of the calibration plate in the image; According to the position information, the pixel distance of the calibration plate in the image is calculated to obtain the calibration factor of the pixel size: Among them, d i is the pixel distance between two adjacent markers, n is the number of pixel distances, D actual It is the actual physical length between two adjacent marking points; Calculate the pixel area of ​​the pavement aggregate area in the image and obtain the pavement aggregate area SA:

Images